Feature gates reference

This page owns the feature-gate taxonomy visible in cli.renamed.js: GrowthBook evaluation, tengu_* gates/signals, environment kill switches, managed-policy gates, and CLI/settings gates that switch operational behavior.

Use Telemetry and tracing for telemetry/export behavior and Environment variables reference for canonical env-var names.

Source anchors

Semantic alias	String or symbol	Meaning
GrowthBookClient	`GrowthBook`	Embedded feature/experiment client.
FeatureFlagLookupWrapper	`Z$("..."`	Feature-flag lookup wrapper pattern.
DisableGrowthBookEnv	`DISABLE_GROWTHBOOK`	Environment kill switch for GrowthBook.
DisableCronEnv	`CLAUDE_CODE_DISABLE_CRON`	Cron/scheduled-task kill switch.
KairosCronGate	`isKairosCronEnabled`	Scheduled-task feature gate wrapper.
CronSchedulerRuntime	`createCronScheduler`	Scheduled-task runtime.
DisableRemoteControlPolicy	`disableRemoteControl`	Managed policy gate for Remote Control.
DisableSkillShellPolicy	`disableSkillShellExecution`	Managed policy gate for skill/slash shell execution.
XaaEnvGate	`CLAUDE_CODE_ENABLE_XAA`	MCP cross-app access is disabled unless this env gate is set.
PluginEvalGate	`tengu_walnut_spire`, `CLAUDE_CODE_WALNUT_SPIRE`	Gates execution of the registered early-access `plugin eval` and `plugin eval init` actions.
RemoteTriggerGate	`tengu_surreal_dali`, `allow_remote_sessions`	Remote routine management is feature-, subscription-, mode-, and policy-gated.
WorkflowGates	`enableWorkflows`, `disableWorkflows`, `CLAUDE_CODE_DISABLE_WORKFLOWS`, `allow_workflows`	User, managed, environment, and organization workflow gates.
SafeModeGate	`--safe-mode`, `CLAUDE_CODE_SAFE_MODE`	Hard-disables customizations for diagnostics.
ScreenReaderGate	`--ax-screen-reader`, `CLAUDE_AX_SCREEN_READER`, `axScreenReader`, `tengu_ax_screen_reader`	CLI/env/settings/feature selection for accessible flat rendering.
ProjectsGates	`allow_projects_tool`, `CLAUDE_PROJECT_UUID`	Requires organization permission and a host-attached Project.
ArtifactGates	`CLAUDE_CODE_DISABLE_ARTIFACT`, `disableArtifact`, `enableArtifact`, `allow_cobalt_plinth`, `tengu_cobalt_plinth`	Combines hard disable, preference, policy, account, and rollout checks.
DesignGates	`allow_design_sync`, `tengu_omelette_fouet`, `tengu_slate_quill`	Controls collaborative design and fixed-schema design-system sync exposure.
ObserverAgentGates	`CLAUDE_CODE_EXPERIMENTAL_OBSERVER_AGENTS`, `tengu_observer_agents_enabled`	Requires explicit process opt-in plus rollout enablement.
AgentTeamsGates	`CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS`, raw argv `--agent-teams`, `tengu_amber_flint`	Requires an explicit process/argv opt-in and a rollout value that is not false.
BuiltInAgentRosterGates	`getBuiltInAgents()`, `CLAUDE_CODE_DISABLE_EXPLORE_PLAN_AGENTS`, `tengu_slate_ibis`, `CLAUDE_AGENT_SDK_DISABLE_BUILTIN_AGENTS`, `disableAgentView`, safe mode, SDK entrypoint	Produces the conditional built-in/native agent roster ~282,731–282,862.
EndConversationGates	`tengu_umber_kestrel`, `allowedEntrypoints`, `modelMeetsEndConversationFloor`	Restricts model-ended conversations by rollout configuration, entrypoint, and model floor.
TeamOnboardingGates	`allow_team_onboarding`, `tengu_flint_harbor_share`	Controls local onboarding workflow policy and the optional hosted share layer.
ComputerUseBoundary	`process.platform === "darwin"`, `--computer-use-mcp`, `computer-use.lock`	Limits native desktop control to the macOS/session-wired path and one controlling session.

Bundle modules in `cli.renamed.js`

Semantic alias	Loader line	Representative renamed exports	Atlas entry
`RemoteControlFeatureGates`	325388	`isRunningInRemoteEnvironment`, `isRemoteControlInternalEventsEnabled`, `isRemoteControlHardDisabled`, `isPreviewHmrEnabled`, `isPersistentRemoteSessionEnabled`, `isCseShimEnabled`, `isCcrV2SendEventsEnabled`, `isCcrMirrorEnabled`	Bundle module map — remote control, feature flags, networking
`GrowthBookFeatureFlags`	167597	`refreshGrowthBookFeatures`, `refreshGrowthBookAfterAuthChange`, `setupPeriodicGrowthBookRefresh`, `stopPeriodicGrowthBookRefresh`, `setGrowthBookConfigOverride`, `onGrowthBookRefresh`, `isGrowthBookEnabled`, `resetGrowthBook`	Bundle module map — remote control, feature flags, networking

Gate taxonomy

flowchart TD
    Decision[Runtime decision] --> GB[GrowthBook / tengu feature flags]
    Decision --> Env[Environment kill switches]
    Decision --> Policy[Managed settings / policy]
    Decision --> Cli[CLI flags]
    Decision --> Settings[User/project/local settings]

    GB --> Behavior[Feature behavior]
    Env --> Behavior
    Policy --> Behavior
    Cli --> Behavior
    Settings --> Behavior
    Behavior --> Evidence[debug / telemetry evidence]

Gate type	Examples	What it controls
GrowthBook/feature flag	`GrowthBook`, `Z$("...")`, many `tengu_*` keys	Experiments, staged rollout, runtime feature defaults.
Environment kill switch	`DISABLE_GROWTHBOOK`, `CLAUDE_CODE_DISABLE_CRON`, telemetry/update/context UI env vars	Coarse process-level enable/disable decisions.
Managed policy/settings	`disableRemoteControl`, `disableAllHooks`, `disableAgentView`, `disableSkillShellExecution`, `cleanupPeriodDays`	Organization/admin and persisted configuration controls.
CLI flag	`--debug`, `--debug-file`, `--add-trace-attribute`, `--no-session-persistence`, `--agents`	Per-invocation behavior.
Provider/env selector	`CLAUDE_CODE_USE_BEDROCK`, `CLAUDE_CODE_USE_VERTEX`, `CLAUDE_CODE_USE_FOUNDRY`, `CLAUDE_CODE_USE_ANTHROPIC_AWS`, `CLAUDE_CODE_USE_MANTLE`	Provider/auth/reporting paths.

High-signal feature-gate families

The table groups source-visible keys by nearby behavior. It does not claim every opaque tengu_* name is fully decoded.

Family	Source-visible gates	Interpreted scope
Remote / bridge / CCR	`tengu_ccr_bridge`, `tengu_remote_backend`, `tengu_bridge_repl_v2_cse_shim_enabled`, `tengu_bridge_attestation_enforce`, `tengu_bridge_attestation_enforce_config`, `tengu_ccr_v2_send_events_cli`, `tengu_bridge_requires_action_details`, `tengu_bridge_system_init`, `tengu_surreal_dali`, `CLAUDE_CODE_REMOTE`, `disableRemoteControl`, `allow_remote_sessions`	Remote Control, bridge transport, attestation, event forwarding, and remote routine management.
Scheduled tasks / Kairos cron	`tengu_kairos_cron`, `tengu_kairos_cron_durable`, `tengu_kairos_loop_dynamic`, `tengu_kairos_loop_persistent`, `tengu_kairos_loop_prompt`, `tengu_kairos_push_notifications`, `tengu_kairos_input_needed_push`, `CLAUDE_CODE_DISABLE_CRON`	Cron/scheduled prompt feature family and autonomous-loop behavior.
Agents / background agents	`tengu_auto_background_agents`, `tengu_agent_list_attach`, `tengu_mcp_subagent_prompt`, `tengu_slim_subagent_claudemd`, `CLAUDE_AGENT_SDK_DISABLE_BUILTIN_AGENTS`, `CLAUDE_CODE_DISABLE_EXPLORE_PLAN_AGENTS`, `tengu_slate_ibis`, `CLAUDE_CODE_DISABLE_AGENT_VIEW`, `CLAUDE_CODE_EXPERIMENTAL_OBSERVER_AGENTS`, `tengu_observer_agents_enabled`, `CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS`, `tengu_amber_flint`	Native roster composition, Agent UI, background agents, subagent prompt/context behavior, experimental observer pairings, and Agent Teams.
Dynamic workflows	`enableWorkflows`, `disableWorkflows`, `workflowKeywordTriggerEnabled`, `CLAUDE_CODE_DISABLE_WORKFLOWS`, `CLAUDE_CODE_WORKFLOWS`, org `allow_workflows`	Workflow availability, keyword opt-in, and policy enforcement.
Hosted projects/artifacts/design/onboarding	`allow_projects_tool`, `CLAUDE_PROJECT_UUID`, `CLAUDE_CODE_DISABLE_ARTIFACT`, `allow_cobalt_plinth`, `tengu_cobalt_plinth`, `allow_design_sync`, `tengu_omelette_fouet`, `tengu_slate_quill`, `allow_team_onboarding`, `tengu_flint_harbor_share`	Attached Project knowledge, hosted page publication, collaborative design, design-system synchronization, and team-guide sharing.
Auto-mode / permission automation	`tengu_auto_mode_config`, `tengu_disable_bypass_permissions_mode`, `tengu_auto_notice_once`, policy `defaultMode=auto`	Auto-mode defaults, consent, permission automation.
Tool/runtime security	`tengu_streaming_tool_execution2`, `tengu_harbor_permissions`, `tengu_destructive_command_warning`, `tengu_bash_allowlist_strip_all`, `CLAUDE_CODE_DISABLE_ADVISOR_TOOL`, `CLAUDE_CODE_ENABLE_EXPERIMENTAL_ADVISOR_TOOL`, `tengu_umber_kestrel`	Tool execution mode, warnings, advisory/permission behavior, and model-ended conversation rollout.
MCP/plugins/skills	`tengu_mcp_directory_visibility`, `tengu_mcp_directory_bff`, `tengu_mcp_singleton_unwrap`, `tengu_skills_dashboard_enabled`, `tengu_plugin_official_mkt_git_fallback`, `CLAUDE_CODE_PLUGIN_PREFER_HTTPS`, `CLAUDE_CODE_PLUGIN_USE_ZIP_CACHE`, `CLAUDE_CODE_SYNC_PLUGIN_INSTALL`, `CLAUDE_CODE_ENABLE_XAA`, `CLAUDE_CODE_ENABLE_REFRESH_MCP_TOOLS`, `tengu_walnut_spire`, `CLAUDE_CODE_WALNUT_SPIRE`	MCP discovery/refresh, plugin install/update/evaluation, skills UI, and cross-app OAuth.
Native computer use	macOS platform guard, internal computer-use startup predicate, `--computer-use-mcp`, session callbacks	Dynamic native desktop-control MCP, separate from Claude-in-Chrome; app approval and lock checks continue after startup.
Context/model behavior	`CLAUDE_CODE_DISABLE_1M_CONTEXT`, `DISABLE_COMPACT`, `DISABLE_INTERLEAVED_THINKING`, `USE_API_CONTEXT_MANAGEMENT`, `CLAUDE_CODE_DISABLE_EXPERIMENTAL_BETAS`, `tengu_prompt_cache_1h_config`, `tengu_prompt_cache_diagnostics`	Context window, compaction, thinking, prompt cache, beta flags.
Web/search/fetch adjacent	`tengu_tool_search_unsupported_models`, `/api/web/domain_info?domain=`	Web/domain metadata and model support guardrails.
UI/terminal behavior	`CLAUDE_CODE_DISABLE_ALTERNATE_SCREEN`, `CLAUDE_CODE_NO_FLICKER`, `CLAUDE_CODE_DISABLE_MOUSE`, `CLAUDE_CODE_DISABLE_MOUSE_CLICKS`, `CLAUDE_CODE_FORCE_SYNC_OUTPUT`, `CLAUDE_AX_SCREEN_READER`, `CLAUDE_CODE_DISABLE_TERMINAL_TITLE`	Terminal rendering, pointer, and screen-reader behavior.
Recovery/corporate launch	`CLAUDE_CODE_SAFE_MODE`, `disableSideloadFlags`, `processWrapper`, `CLAUDE_CODE_PROCESS_WRAPPER`	Configuration isolation, extension sideload policy, and required launch wrappers.
Updates	`DISABLE_UPDATES`, `DISABLE_AUTOUPDATER`, `FORCE_AUTOUPDATE_PLUGINS`, `auto-update` settings	Native/plugin updater behavior.
Telemetry/diagnostics	`CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC`, `DISABLE_TELEMETRY`, `DO_NOT_TRACK`, `DISABLE_ERROR_REPORTING`, `DISABLE_GROWTHBOOK`, `OTEL_LOG_TOOL_DETAILS`, `OTEL_LOG_TOOL_CONTENT`, `OTEL_LOG_USER_PROMPTS`	Observability, analytics, error reporting, feature-evaluation fetching.

Relationship to sessions and agents

flowchart TD
    Gates[Feature/env/policy gates] --> Session[Session envelope]
    Gates --> Agents[Agent/task scheduler]
    Gates --> Tools[Tool/security runtime]
    Gates --> Telemetry[Telemetry and tracing]
    Session --> Evidence[debug/telemetry evidence]
    Agents --> Evidence
    Tools --> Evidence

Runtime area	Gate interaction
Sessions/remote	Remote Control can be disabled by managed policy; bridge state and transcript mirrors emit observable frames.
Tools/permissions	Permission automation and streaming execution are feature-gated; denials can surface as stream frames and telemetry.
Agents/tasks	Background agents, Agent Teams, observer agents, subagent prompt shape, scheduled tasks, and auto-mode are behind feature/env/policy gates.
Conversation lifetime	`EndConversation` requires the rollout config, entrypoint regex, and supported model; its two-call reflection and durable marker are additional runtime guards.
Scheduled tasks	`CLAUDE_CODE_DISABLE_CRON` disables cron; scheduled-task telemetry tracks missed/fire/expired behavior.
Hosted review/update	Hosted review preflight and native updater emit operational telemetry but do not own model turns.

Safe interpretation rules

Treat tengu_* as evidence of a gate or telemetry event, not a fully decoded feature name unless adjacent strings confirm behavior.
Prefer env/policy strings for hard claims, because they are usually descriptive and user/admin-facing.
Distinguish a gate from an exported event: feature gates decide behavior; telemetry/tracing reports behavior.
Use Telemetry and tracing for sink/export details.

Focused lifecycle gates

Lifecycle	Gate chain	Important distinction
Observer agents	`CLAUDE_CODE_EXPERIMENTAL_OBSERVER_AGENTS` must be true; `tengu_observer_agents_enabled` must not disable the feature; the named observer type must exist and cannot itself declare an observer.	Even after feature enablement, arm and every digest delivery re-run Agent permission and managed `PreToolUse` checks.
Agent Teams	`CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS` or the raw argv token `--agent-teams`, plus `tengu_amber_flint !== false`; startup must be an interactive non-teammate lead to initialize the implicit team.	`teammateMode` selects execution only after enablement; disk-backed task/mailbox files do not turn the clean-exit-deleted team into a durable service.
Built-in/native agents	Normal mode seeds `general-purpose`; safe mode controls `statusline-setup`; Agent-view policy controls `claude`; `CLAUDE_CODE_DISABLE_EXPLORE_PLAN_AGENTS`/`tengu_slate_ibis` control `Explore` + `Plan`; SDK entrypoint controls `claude-code-guide`; coordinator mode substitutes `worker`.	“Built-in” is provenance, not universal visibility. Nine source definitions do not mean nine simultaneously selectable personas.
Conversation termination	`modelMeetsEndConversationFloor` + `tengu_umber_kestrel` + configured/default entrypoint regex + runtime exclusion check.	Visibility does not end a conversation: the same tool must be called across two assistant turns before the marker/abort path runs.
Team onboarding and share flows	`allow_team_onboarding` controls the user-only workflow. Hosted sharing additionally requires nonessential traffic, stored OAuth, and `tengu_flint_harbor_share`.	A local `ONBOARDING.md` remains useful when hosted sharing is unavailable.
Computer-use MCP	macOS + interactive/session-wired startup + internal enablement predicate.	Startup visibility is only the first boundary; TCC, per-app approval/tier, optional grants, live target checks, and a process lock govern calls.
Plugin evaluation	`tengu_walnut_spire` or `CLAUDE_CODE_WALNUT_SPIRE` must pass the action-time guard; normal case/tool/model/cost controls apply afterward.	The gate controls a CLI evaluation harness, not an ordinary plugin contribution loaded into every session.

GrowthBook client lifecycle

The GrowthBookFeatureFlags module (loader at cli.renamed.js:167597, body at cli.renamed.js:167263) is the single source of truth for tengu_* features. It wraps a GrowthBook SDK client (wS$) initialized lazily against https://api.anthropic.com/. The module exports three read APIs with different freshness/blocking trade-offs, plus periodic refresh, auth-change reset, and config-override helpers.

Initialization (`initializeGrowthBook`, `Qq6`)

Both factories are memoized via L8(...):

Disable kill switch — isGrowthBookEnabled() returns false when DISABLE_GROWTHBOOK env is set or DF() (a baseline gate) is off. All read APIs degrade to the caller’s defaultValue in this case.
Trust gate — auth headers are only attached when one of checkHasTrustDialogAccepted(), getSessionTrustAccepted(), or getIsNonInteractiveSession() holds; otherwise the client is constructed without apiHostRequestHeaders and the initial init({timeout: 5000}) runs unauthenticated (and gets no remote-eval response). The flag gq6 records whether auth was attached.
Client construction — passes apiHost, clientKey (per-build), attributes (see getUserAttributes()), remoteEval: true, and cacheKeyAttributes: ["id", "organizationUUID"]. Remote evaluation means the server returns only the feature values for this user’s attributes, so the local cache never contains other users’ assignments.
init({timeout: 5000}) — the resulting promise is stored in u5$. checkSecurityRestrictionGate(...) awaits u5$ so its first call after auth-change blocks until features are loaded.
FTK(client) — extracts the features payload, separates experiment-sourced features (tracked in sYH), normalizes value/defaultValue, populates the in-memory cache jF, and clears stale state. Returns true if anything changed.
Process hooks — registers beforeExit / exit listeners (C5$, b5$) that destroy the client on shutdown.
Periodic refresh — setupPeriodicGrowthBookRefresh() schedules a setInterval(refreshGrowthBookFeatures, 21600000) (6 h) and registers a beforeExit listener to clean it up.

User attributes (`getUserAttributes`)

The attribute bag used for GrowthBook targeting and cacheKeyAttributes:

Attribute	Source
`id`, `deviceID`, `sessionId`, `platform`	Stable device + session identity.
`organizationUUID`, `accountUUID`	OAuth account + `CLAUDE_CODE_ORGANIZATION_UUID` / `CLAUDE_CODE_ACCOUNT_UUID` env overrides.
`apiBaseUrlHost`	Non-default `ANTHROPIC_BASE_URL` host.
`userType`, `subscriptionType`, `rateLimitTier`, `firstTokenTime`, `email`, `appVersion`	Identity service / config.
`githubActionsMetadata`	Present when running inside GitHub Actions.
`releaseChannel`	Auto-update channel from initial settings.
`entrypoint`	`NRH()` — startup entry classification.

Only id + organizationUUID are part of cacheKeyAttributes, so cache invalidation is per device per org.

Read APIs and cache tiers

API	Blocks?	Uses in-memory `jF`?	Uses persisted `cachedGrowthBookFeatures`?	Tracks experiment exposure?
`getFeatureValue_DEPRECATED(key, default)`	Yes — awaits init.	Yes (preferred).	Falls back to remote eval.	Yes (`Ri$(key)`).
`getFeatureValue_CACHED_MAY_BE_STALE(key, default)`	No.	Yes (preferred).	Yes (fallback).	Only if `sYH.has(key)` (already known experiment). Otherwise records `x5$` for batched tracking after next refresh.
`getFeatureValue_CACHED_WITH_REFRESH(key, default, _)`	No (alias of the stale getter; the refresh side effect lives elsewhere).	Yes.	Yes.	Same as stale.
`checkGate_CACHED_OR_BLOCKING(key)`	Only when not cached AND not env-/override-set.	Fast path: `cachedGrowthBookFeatures[key] === true` returns immediately.	Yes.	Yes on fast path.
`checkSecurityRestrictionGate(key)`	Yes — awaits in-flight `u5$` to ensure post-auth-change correctness; then reads `cachedGrowthBookFeatures`.	No.	Yes.	No (security gates do not feed experiment tracking).
`getDynamicConfig_BLOCKS_ON_INIT(key, default)` / `getDynamicConfig_CACHED_MAY_BE_STALE(key, default)`	Aliases of the deprecated/stale getters for typed dynamic-config blobs.	—	—	—

Two override layers run before any of these:

Env override — m5$() lazily reads BTK (parsed JSON from a hosted env). Any key in there wins.
Config override — B5$() reads a runtime override map (currently always undefined for end users — the setter is a no-op).

Caching surfaces

jF: Map<string, value> — in-memory cache populated by FTK(client).
sYH: Map<string, {experimentId, variationId}> — features whose source is an experiment; experiment exposure is tracked once per process via Ri$(key).
Si$: Set<string> — feature keys whose value is not the default (getNonDefaultFeatureKeys()).
x5$: Set<string> — keys observed by _CACHED_MAY_BE_STALE before init completed; the post-init init().then(...) callback walks this set and emits experiment exposure for any keys that turn out to be experiments.
Uq6: Set<string> — keys already tracked, used to dedupe exposure events.
cachedGrowthBookFeatures / cachedExperimentFeatures — persisted to global config via QTK() after each successful FTK. This is what _CACHED_MAY_BE_STALE reads when jF is empty (process startup) so the very first cold render still has a cached value.

Refresh and reset

Function	Trigger	Behavior
`setupPeriodicGrowthBookRefresh()`	After first init.	`setInterval(refreshGrowthBookFeatures, 21600000)` (6 h). `unref()`’d so it does not keep the event loop alive.
`stopPeriodicGrowthBookRefresh()`	`beforeExit`, manual reset.	Clears the interval and removes the `beforeExit` listener.
`refreshGrowthBookFeatures()`	Periodic timer, manual.	Calls `client.refreshFeatures({skipCache: true})`, then `FTK(client)`. On change: persists via `QTK()` and emits `imH.emit()` so subscribers (`onGrowthBookRefresh`) re-evaluate.
`refreshGrowthBookAfterAuthChange()`	Called by auth flows after login/logout/token refresh.	`resetGrowthBook()` then re-runs `initializeGrowthBook()` and tracks the in-flight promise in `u5$` so blocking gates await the new client.
`resetGrowthBook()`	Auth change, shutdown.	Stops the interval, removes process listeners, destroys the client, clears every cache map (`jF`, `sYH`, `Si$`, `x5$`, `Uq6`), and invalidates the memoized factories.
`onGrowthBookRefresh(cb)`	UI/runtime subscribers.	Subscribes `cb` to `imH`. If the cache already has values (`jF.size > 0`), schedules an immediate microtask call so late subscribers still see current state. Returns an unsubscribe handle.

Remote Control / Bridge feature gates

The RemoteControlFeatureGates module (loader at cli.renamed.js:325388, body at cli.renamed.js:325146) layers business logic on top of GrowthBook to decide whether Claude Code’s Remote Control / Bridge / CCR / CSE features are available in the current session.

Bridge entitlement chain

flowchart TD
    Start[isBridgeEnabled / Blocking] --> A{TX6 dev override?}
    A -->|yes| Enable[enabled]
    A -->|no| B{isRemoteControlHardDisabled?}
    B -->|yes| Disable[disabled]
    B -->|no| C{isRunningInRemoteEnvironment?}
    C -->|yes| Disable
    C -->|no| D{hj API==Anthropic?}
    D -->|no| Disable
    D -->|yes| E{$78 claude.ai subscriber?}
    E -->|no| Disable
    E -->|yes| F{tengu_ccr_bridge feature?}
    F -->|cached miss| G[isBridgeEnabledBlocking awaits checkGate_CACHED_OR_BLOCKING]
    F -->|cached true| Enable
    G -->|true| Enable
    G -->|false| Disable

Functions:

isRemoteControlHardDisabled() — managed setting disableRemoteControl: true.
isRunningInRemoteEnvironment() — CLAUDE_CODE_REMOTE env OR getIsRemoteMode(). Bridge is one-way: a Bridge process never enables Bridge for itself.
hasBridgeEntitlement() — synchronous, uses _CACHED_MAY_BE_STALE. Combines hj() + claude.ai subscriber check + tengu_ccr_bridge feature value.
isBridgeEnabled() — fast-path; calls hasBridgeEntitlement and the hard-disable checks.
isBridgeEnabledBlocking() — async; uses checkGate_CACHED_OR_BLOCKING("tengu_ccr_bridge") so the first call after auth change correctly waits for refreshed features.

`getBridgeDisabledReason()` chain

When Bridge is unavailable the UI calls getBridgeDisabledReason() to render a user-friendly explanation. It returns the first matching reason:

API host is not api.anthropic.com ⇒ “only available when using Claude via api.anthropic.com”.
isRunningInRemoteEnvironment() ⇒ “not available inside a remote session”.
isRemoteControlHardDisabled() ⇒ “disabled by your organization’s policy (managed setting disableRemoteControl)”.
Not a claude.ai subscriber ⇒ “requires a claude.ai subscription. Run claude auth login…”.
Auth-helper conflict (Y2_()) ⇒ describes the conflicting API key / token source (ANTHROPIC_API_KEY, apiKeyHelper, ANTHROPIC_AUTH_TOKEN, ANTHROPIC_UNIX_SOCKET, etc.) and how to unset it.
No full-scope profile token (e.g. only claude setup-token long-lived token) ⇒ “Long-lived tokens are limited to inference-only…”.
Missing oauthAccount.organizationUuid ⇒ “Unable to determine your organization…”.
checkGate_CACHED_OR_BLOCKING("tengu_ccr_bridge") returns false ⇒ “not yet enabled for your account”.
Otherwise null (Bridge is available).

getBridgeAuthDebugInfo() — only emits content when vv() (a debug-flag predicate) is true. Returns a multi-line block enumerating isBareMode, OAuth token presence and scopes, claude.ai inference scope, claude.ai subscriber flag, profile scope, org UUID, plus the relevant env vars (ANTHROPIC_API_KEY, ANTHROPIC_AUTH_TOKEN, apiKeyHelper, CLAUDE_CODE_API_KEY_FILE_DESCRIPTOR, CLAUDE_CODE_OAUTH_TOKEN, ANTHROPIC_UNIX_SOCKET, and any third-party provider env (USE_BEDROCK, USE_VERTEX, USE_FOUNDRY, USE_ANTHROPIC_AWS, USE_MANTLE)).

Secondary feature gates

Function	Backing GrowthBook key	Default	Effect
`isCseShimEnabled()`	`tengu_bridge_repl_v2_cse_shim_enabled`	`true`	Bridge REPL v2 CSE shim. Default true means turning the feature off requires an explicit flip.
`isPreviewHmrEnabled()`	`tengu_bridge_vivid`	`false`	Preview HMR mode for Bridge.
`isRemoteControlInternalEventsEnabled()`	`tengu_amber_relay`	`false`	Internal-event relay for Remote Control.
`isCcrV2SendEventsEnabled()`	`tengu_ccr_v2_send_events_cli`	`false`	Whether CCR v2 sends event frames.
`isCcrMirrorEnabled()`	(hard-coded `false`)	—	Reserved; currently always disabled.
`isPersistentRemoteSessionEnabled()`	(hard-coded `false`)	—	Reserved; currently always disabled.
`getCcrAutoConnectDefault()`	derived	`false`	True only when persistent-remote is enabled AND we are not in a remote env.
`getAttestationFilterPolicy()`	`tengu_bridge_attestation_enforce` (+ `tengu_bridge_attestation_enforce_config`)	enforce off	When enforcement is on, returns a parsed policy object via `BM7(...)`; otherwise returns the default `ZX6`.
`checkBridgeMinVersion()`	`tengu_bridge_min_version` dynamic config	`0.0.0`	Compares the bundle `VERSION` (`2.1.215` for this build) against the required minVersion. Returns the formatted error message if too old, else `null`.

App-state appliers

applyRemoteControlToAppState(state, enabled) — toggles replBridgeEnabled and clears replBridgeOutboundOnly. Short-circuits when an outbound-only session is already in progress.
applyAutoUploadSessionsToAppState(state, enabled) — when neither flag would change, returns the same state; otherwise sets both replBridgeEnabled and replBridgeOutboundOnly to the new value. This is how the daemon-side “auto-upload sessions” toggle quietly opens an outbound-only Bridge connection.

GrowthBook ↔ Remote-Control linkage

The complete dependency chain for “should this session have Remote Control?”:

flowchart LR
    AuthChange[Auth login/logout/refresh] --> Reset[refreshGrowthBookAfterAuthChange]
    Reset --> Init[initializeGrowthBook → wS$ client]
    Init --> Refresh[refreshFeatures + FTK populates jF]
    Refresh --> Persist[cachedGrowthBookFeatures saved to global config]
    Refresh --> Emit[imH.emit triggers subscribers]
    Emit --> RC[Remote-Control gates re-evaluate]
    RC --> Gate[isBridgeEnabled / isCseShimEnabled / isPreviewHmrEnabled / ...]
    Periodic[setInterval 6h] --> Refresh
    Override[BTK env override / B5$ runtime override] -. wins over .-> Gate

The practical outcomes:

Logging in / out flips Bridge entitlement immediately because refreshGrowthBookAfterAuthChange resets the client and the next _CACHED_MAY_BE_STALE read will block on the in-flight init via checkSecurityRestrictionGate.
Operators can pin a build’s behavior via env overrides (parsed once by m5$()); those win over both the in-memory cache and the persisted cache.
The 6-hour periodic refresh is unref’d, so it never keeps the process alive past its natural exit; daemon-mode processes still get periodic updates because they have their own keep-alive sources.

Created and maintained by Yingting Huang.