openhuman/.claude/memory.md

Project Memory

Quick reference for anyone starting with Claude on this project. Updated by the memory-keeper agent.

Fixes & Gotchas

• macOS close button does not dismiss window (issue #2049) — WebviewWindow::hide() routes through CEF's WindowMessage::Hide → cef::Window::hide() which does NOT propagate to the visible NSWindow frame. Fix: use AppHandle::hide() which calls [NSApp hide:] via set_application_visibility(false). This is macOS-only (#[cfg(target_os = "macos")]); the CloseRequested handlers are in app/src-tauri/src/lib.rs (grep WindowEvent::CloseRequested — line numbers drift). PR #2118.
• ServiceBlockingGate CORS errors — The gate calls openhumanServiceStatus() and openhumanAgentServerStatus() at startup. These used callCoreRpc() which falls back to raw fetch() when socket isn't connected yet, causing CORS errors. Fix: route through invoke('core_rpc_relay') instead (Tauri IPC, no CORS).
• Socket not connected at startup — SocketProvider only connects when a Redux auth.token is set. At fresh launch (no token), socket is null, so any callCoreRpc() call falls back to fetch(). Always use invoke('core_rpc_relay') for local sidecar RPC calls.
• openhuman.agent_server_status doesn't exist — This RPC method is not registered in the core. The gate checks it but it always errors. The gate passes if either service is Running OR agent server is running OR core is reachable.
• Cargo incremental builds can serve stale UI — If the app shows old frontend after a Rust rebuild, run cargo clean --manifest-path app/src-tauri/Cargo.toml before rebuilding.
• build.rs missing rerun-if-changed causes stale ACL / "Command not found" at runtime — app/src-tauri/build.rs had no cargo:rerun-if-changed directives for permissions/ or capabilities/. Adding/changing TOML or JSON files there did not re-trigger tauri-build, so ACL tables were stale and registered commands silently failed. Fixed by adding println!("cargo:rerun-if-changed=permissions") and println!("cargo:rerun-if-changed=capabilities") in build.rs (issue #270). Also: any new Tauri command must have a matching entry in a permissions/ TOML file or it will hit the same error even if it is in generate_handler!.
• macOS deep links require .app bundle — pnpm tauri dev does NOT support deep links. Must use pnpm tauri build --debug --bundles app.

Strict Rules

• No dynamic imports in app/src/ — Use static import at file top. Guard call sites with try/catch for Tauri/non-Tauri safety. See CLAUDE.md.
• Service RPC calls must use Tauri IPC — Never use callCoreRpc() for service operations. Use invoke('core_rpc_relay', { request: { method, params } }).
• All frontend env vars go through app/src/utils/config.ts — Never read import.meta.env.VITE_* directly in other files. Import from config.ts instead. See .env.example files for the full list.
• Always run checks before commit — pnpm workspace openhuman-app compile, pnpm lint, pnpm format:check, pnpm build, pnpm tauri dev. Husky hooks enforce some but run all manually first.
• Stage specific files — Never git add -A. Always git add <specific-files>.

Workflow

• Agent order: architectobot (plan) → user approval → codecrusher (implement) → architectobot (verify)
• Always read CLAUDE.md first before any issue work
• Ask user when in doubt — never assume scope or approach
• PRs target upstream — tinyhumansai/openhuman main branch, not fork
• GraphQL project board can return empty — gh project item-list on board #2 sometimes returns no items even when issues exist. Fall back to gh issue list --repo tinyhumansai/openhuman directly.
• jq regex: use POSIX classes, not \s — jq's test() uses ONIG regex; \s is not supported. Use [[:space:]] for whitespace matching in gh pr list --json ... --jq pipelines.
• PR conflict check: Closes #N syntax not always used — gh pr list --jq "select(.body | test('Closes #N'))" misses PRs that mention an issue thematically without a closing keyword. Also search PR title + body for the raw issue number (#N) with broader matching to catch related open PRs before claiming an issue is unassigned.
• pnpm debug unit path is relative to app/src/ — Pass providers/__tests__/Foo.test.tsx, not app/src/providers/__tests__/Foo.test.tsx.
• Prettier must run after codecrusher adds test cases — New test blocks often fail format:check. Run pnpm --filter openhuman-app format before committing when test files are touched.
• Check for existing PRs before implementing — When the workflow picks an issue, search open PRs for the issue number and related keywords before starting work. A contributor may have already shipped the fix (e.g. PR #2101 for issue #2075).
• Project board gh project item-list paginates closed items first — The first 100 items returned are often CLOSED. Must --limit 500 or paginate to find open/unassigned work. Fall back to gh issue list --repo tinyhumansai/openhuman --state open for reliability.

Local AI Presets

• Tier system lives in src/openhuman/local_ai/presets.rs — single source of truth for tier→model ID mapping. To change default models for a release, edit all_presets() there.
• Device detection uses sysinfo crate (src/openhuman/local_ai/device.rs). Apple Silicon = GPU always; others = best-effort.
• OPENHUMAN_LOCAL_AI_TIER env var overrides the selected tier at config load time (in load.rs).
• Frontend tier selector is in LocalModelPanel.tsx under Settings > Local AI Model. Uses coreRpcClient to call 3 RPC methods: local_ai_device_profile, local_ai_presets, local_ai_apply_preset.
• Default config maps to Medium tier (gemma3:4b-it-qat). If someone changes model_ids.rs defaults, they should keep presets.rs in sync.
• ollama_base_url() previously ignored config.local_ai.base_url — It only read env vars. Fixed in feat/ollama-external-server-url by adding ollama_base_url_from_config(config). Any new Ollama URL resolution must go through the config-aware helper, not the env-only one.
• LocalModelDebugPanel.tsx must seed URL from config on mount — Previously initialized ollamaBaseUrlInput to the hardcoded default and only loaded the persisted URL when diagnostics ran. Fix: useEffect on mount calls openhumanGetConfig() and sets state from config.local_ai.base_url. Pattern to follow for any settings field backed by Rust config.

Core process (in-process, no sidecar)

• Core runs in-process as a tokio task inside the Tauri host (sidecar removed in PR #1061). Lifecycle owned by core_process::CoreProcessHandle in app/src-tauri/src/core_process.rs.
• pnpm core:stage is a no-op echo — there is no target/debug/openhuman-core binary to copy into app/src-tauri/binaries/. Rebuilding via pnpm dev:app is enough to pick up new RPC methods in the in-process server.
• Token auth: per-launch hex bearer in OPENHUMAN_CORE_TOKEN, exposed to the renderer via core_rpc_token Tauri command. Always call RPC through invoke('core_rpc_relay', { request: { method, params } }) — avoids the CORS preflight fetch() would trigger.
• Stale-listener policy (#1130): if the port is already in use, the handle probes GET / to decide if it's an OpenHuman core, then term/force-kill with PID revalidation guarding against PID reuse. Set OPENHUMAN_CORE_REUSE_EXISTING=1 to attach to a manually-started openhuman-core serve for debugging.
• Default port 7788. Stage token (when running standalone): ~/.openhuman-staging/core.token under OPENHUMAN_APP_ENV=staging.

Onboarding System

• OnboardingOverlay is a portal, not a route — mounted in App.tsx, renders via createPortal at z-[9999]. There is no /onboarding route in AppRoutes.tsx. Gating is purely Redux + workspace flag.
• Deferred onboarding — onboardingDeferredByUser in authSlice.ts (persisted via redux-persist) durably tracks when a user clicks "Set up later". SetupBanner.tsx provides the resume path.
• selectHasIncompleteOnboarding is unused in production code — only tested. Don't use it for new features.
• Logout must clear onboarding state — _clearToken resets isOnboardedByUser + isAnalyticsEnabledByUser. Workspace flag (.skip_onboarding file) is cleared via openhumanWorkspaceOnboardingFlagSet(false) in SettingsHome logout, clearAllAppData, and UserProvider auth recovery. All three paths must stay in sync. OnboardingOverlay local state (userLoadTimedOut, onboardingCompleted) is reset via a useEffect watching token — if token becomes null, both reset to initial values (#192).
• LocalAI download errors must surface — LocalAIStep has an onDownloadError callback prop; Onboarding.tsx renders an error banner via createPortal when it fires. Without this, download failures are silently swallowed (#194).
• formatBytes / formatEta / progressFromStatus — shared in app/src/utils/localAiHelpers.ts. Home.tsx and LocalModelPanel.tsx still have local copies (can be migrated later).
• Notification z-index stacking — ErrorReportNotification: z-[10000] bottom-right. OnboardingOverlay: z-[9999]. LocalAIDownloadSnackbar: z-[9998] bottom-left.
• React Compiler lint — useCallback deps must match the full inferred closure. Using user?._id as dep when the closure captures user triggers preserve-manual-memoization. Use user as the dep instead.
• setState in effects — ESLint react-hooks/set-state-in-effect catches synchronous setState in useEffect bodies. Use lazy initializers, compute at render, or event handlers instead.
• Walkthrough is multi-page (9 steps) — Uses react-joyride v3 Step.before async hooks to navigate between pages (/home → /chat → /skills → /intelligence → /settings → /home). Steps factory: createWalkthroughSteps(navigate) in walkthroughSteps.ts. waitForTarget(selector, timeout) polls via rAF until DOM target appears. Re-trigger from Settings via resetWalkthrough() + walkthrough:restart CustomEvent. AppWalkthrough is mounted inside Router context (can use useNavigate directly). BottomTabBar attr is tab-notifications (not tab-automation).
• OnboardingNextButton is the shared primary CTA — All onboarding steps use app/src/pages/onboarding/components/OnboardingNextButton.tsx. New steps must use this component for the primary navigation button.
• Onboarding is 3 steps: Welcome(0) → Skills(1) → ContextGathering(2) — Referral step was removed (issue #752). ReferralApplyStep.tsx is preserved but unused. referralApi is still used on the Rewards page. WelcomeStep no longer has nextDisabled/nextLoading/nextLoadingLabel props (those gated on referral stats prefetch).
• Recovery Phrase moved to Settings — MnemonicStep was removed from onboarding (was step 5). The same BIP39 generate/import functionality now lives in app/src/components/settings/panels/RecoveryPhrasePanel.tsx, accessible via Settings > Recovery Phrase. Onboarding completion logic moved into handleSkillsNext in Onboarding.tsx.
• E2E tests find onboarding buttons by label text — shared-flows.ts, login-flow.spec.ts, auth-access-control.spec.ts, and voice-mode.spec.ts locate buttons by their visible label. Changing button labels requires updating all four files. Note: voice-mode.spec.ts still references legacy labels that don't match current steps (pre-existing tech debt).
• ScreenPermissionsStep always shows Continue — The Continue button is always visible regardless of permission grant status, allowing users to skip the permissions step (#274).
• OnboardingOverlay RPC/Redux race condition — getOnboardingCompleted() RPC can fail (sidecar not ready, timeout); the old catch block hardcoded setOnboardingCompleted(false), ignoring the persisted isOnboardedByUser Redux flag. Fix: read selectIsOnboarded from authSelectors.ts in the catch block as fallback, and combine both flags in shouldShow: !onboardingCompleted && !isOnboardedRedux. Either flag being true is sufficient to skip onboarding (#197).
• DEV_FORCE_ONBOARDING was a no-op — The old ternary had identical branches; fixed to actually force-show when the flag is set.
• isOnboardedRedux must be in useEffect deps — When reading a selector value inside a useEffect, add it to the dependency array or the effect won't re-run when Redux state changes.

CoreStateProvider & Auth Bootstrap

• Auth session tokens are NOT in Redux persist — They live entirely in the Rust sidecar, fetched via fetchCoreAppSnapshot() RPC. PersistGate only gates non-auth state (AI config, threads, channel connections). CoreStateProvider bootstrap is the critical auth path.
• CoreStateProvider premature isBootstrapping: false causes blank Settings — If the initial RPC call fails (sidecar still starting), the old error handler set isBootstrapping: false immediately, causing ProtectedRoute to redirect to / before the 3s poll could recover. Fix (issue #413): keep isBootstrapping: true on initial failure, let the poll retry, give up after 5 attempts (~15s).
• CoreStateProvider is consumed by ~25 components — Changes to its state shape or bootstrap behavior affect routes, socket, onboarding, nav, settings, and hooks. Treat it as a high-blast-radius file.
• bootstrapFailCountRef retry counter bug (issue #2158) — The ref is a cumulative lifetime counter; logging it against MAX_BOOTSTRAP_RETRIES (5) as denominator produced impossible attempt 11/5. Fix: distinguish bootstrap phase ("attempt X/5") from continuous-poll phase (separate message, 10s backoff). Reset the counter to 0 on any successful snapshot fetch.
• Settings is a full route, not a modal — /settings/* uses nested <Routes> in Settings.tsx. The .claude/rules/15-settings-modal-system.md doc describing a portal/modal approach is outdated. A catch-all <Route path="*"> redirects unmatched sub-paths to /settings.
• PersistGate loading={null} causes flash — Changed to loading={<RouteLoadingScreen />} (issue #413). RouteLoadingScreen accepts an optional label prop (defaults to "Initializing OpenHuman...") and can be rendered with no props.

Build Blockers: macOS Tahoe + whisper-rs

• whisper-rs breaks cargo build on macOS Tahoe (Apple Silicon) — Added in main via whisper-rs = "0.16" (voice feature #178). Apple clang 21+ refuses -mcpu=native when --target=arm64-apple-macosx is also set. This is NOT fixable by updating CLT.
• Root cause — ggml cmake sets GGML_NATIVE=ON by default; the cmake crate appends --target to clang, triggering the incompatibility. Happens even with the latest toolchain.
• Workaround — Patch ~/.cargo/registry/src/index.crates.io-*/whisper-rs-sys-0.15.0/build.rs: add config.define("GGML_NATIVE", "OFF"); (for target_os = "macos" && target_arch = "aarch64") just before the config.build() call.
• Patch is fragile — Resets on cargo clean, crate version bump, or registry re-download. Deleting build cache alone (target/debug/build/whisper-rs-sys-*) is NOT enough — cmake regenerates with the same bad flags.
• Correct fix — Needs an upstream patch in whisper-rs-sys or a Cargo feature to opt out of GGML_NATIVE on Apple Silicon cross-builds.

UI Redesign (Light Theme — April 2026)

• Full dark-to-light redesign shipped — All pages, components, and settings panels converted from dark glass-morphism to clean light theme based on Figma designs by Mithil (OpenHuman-Prod file, node 2094-250136 for tokens).
• Design tokens saved in my_docs/figma-design-tokens.md — neutral grayscale, primary blue #2F6EF4, success #34C759, alert #E8A728, error #EF4444, SF Pro typography scale.
• Navigation changed: Left MiniSidebar → bottom BottomTabBar (Home, Chat, Skills, Intelligence, Automation, Notification). Settings accessible via gear icon on Home page header.
• MiniSidebar.tsx retained (not deleted) as backup. BottomTabBar.tsx is the active nav component.
• Agent message bubbles need bg-stone-200/80 (not bg-stone-100) on #F5F5F5 background — bg-stone-100 is nearly invisible.
• ~55 files touched — purely CSS class changes, zero logic/handler/state changes.

Upsell / Billing (Phase 1 — Issue #403)

• Upsell components live in app/src/components/upsell/ — UpsellBanner, UsageLimitModal, GlobalUpsellBanner, upsellDismissState. Shared hook: app/src/hooks/useUsageState.ts.
• Usage data sources — creditsApi.getTeamUsage() returns TeamUsage (rolling 10h spend/cap + weekly budget/remaining). billingApi.getCurrentPlan() returns CurrentPlanData (plan tier, caps, subscription status). Both go through callCoreCommand (core RPC). No Redux slice — all local hook state.
• Module-level cache in useUsageState — _cache variable with 60s TTL prevents duplicate API calls when multiple components mount simultaneously. New pattern; do not remove.
• Banner dismiss state uses localStorage (prefix openhuman:upsell:), not Redux — consistent with CLAUDE.md exception for ephemeral UI state.
• Phased rollout — Phase 1 = banners + limit modal + hook. Phase 2 = onboarding upsell + analytics. Phase 3 = remote config + A/B testing.
• "5-hour" label stragglers in Conversations.tsx — LimitPill label and its hover tooltip still say "5h" / "5-hour". Commit 8c52236's "10-hour" terminology refactor missed those two spots.
• getTeamUsage() now normalizes via normalizeTeamUsage() — Added in issue #482. The Rust sidecar passes backend JSON through opaquely (src/openhuman/team/ops.rs), so the TS client must normalize field names and types. Pattern matches existing normalizeCreditBalance() in the same file. Any new billing API that returns raw backend data should follow the same normalize-at-the-client pattern.
• Two separate TeamUsage types exist — creditsApi.ts:24 (billing: cycle budget, limits) and types/team.ts:11 (team model: daily token limit). Different import paths, no collision, but confusing.

Settings & Skills Reorganization (Issue #396)

• Settings is NOT a modal — It's a full route (/settings/*) with nested <Routes>. The .claude/rules/15-settings-modal-system.md doc is outdated.
• SettingsHeader breadcrumbs — All panels now receive breadcrumbs from useSettingsNavigation() hook. The hook derives breadcrumbs from the current route path. When adding a new settings panel, destructure breadcrumbs from the hook and pass to <SettingsHeader>.
• Standard settings padding — All settings panel content areas use p-4 space-y-4. Don't deviate.
• Dead code removed — TauriCommandsPanel, useSettingsAnimation, SettingsPanelLayout, SettingsBackButton, ProfilePanel, AdvancedPanel, SkillsPanel, SkillsGrid were all deleted. Don't re-create them.
• Skills page is the single management surface — Browser Access toggle moved from SkillsPanel to the Skills page. There is no /settings/skills route anymore.
• Panel decomposition — LocalModelPanel, AutocompletePanel, CronJobsPanel, ScreenIntelligencePanel were split into sub-components in subdirectories. Each orchestrator is ≤ ~300 lines.
• UnifiedSkillCard — All skill types (built-in, channels, 3rd party) use UnifiedSkillCard from app/src/components/skills/SkillCard.tsx. Secondary actions use an overflow menu. data-testid attributes (skill-sync-button-*, skill-debug-button-*) must be preserved.
• SkillSearchBar + SkillCategoryFilter — New components in app/src/components/skills/ for search and category filtering on the Skills page.

Composio Backend URL Bug (Issue #2075, PR #2101)

• effective_backend_api_url env-fallback branch skipped normalization — In src/api/config.rs, the override branch normalized via normalize_backend_api_base_url but the env-fallback branch (OPENHUMAN_BACKEND_API_URL) did not, so scheme-less URLs like api.example.com were used raw. Fix: normalize the env-fallback branch too (3-layer defense: config → env-fallback → IntegrationClient::new).
• normalize_backend_api_base_url and redact_url_for_log are pub(crate) — Available for reuse across src/api/ after PR #2101 merge.

Composio Identity (Issue #691)

• ProviderUserProfile.profile_url — New optional field on the struct in src/openhuman/composio/providers/types.rs. Providers should populate it when available from upstream profile payloads.
• identity_set callback in default flow — ComposioProvider::on_connection_created() in src/openhuman/composio/providers/traits.rs now calls identity_set(&profile) after profile fetch. composio_get_user_profile in src/openhuman/composio/ops.rs also routes persistence through identity_set.
• Facet key format for connected identities — skill:{toolkit}:{identifier}:{field} (e.g. skill:gmail:user@example.com:profile_url). Use FacetType::Skill when storing. Toolkit and identifier together form the unique identity; field is the attribute name.
• Connected identities loader/renderer — src/openhuman/composio/providers/profile.rs contains load_connected_identities() (reads skill:* facets) and render_connected_identities_section() (formats markdown for prompt injection). Keep rendering logic there, not in prompt modules.
• Prompt injection helper — render_connected_identities is imported and called in welcome/prompt.rs, orchestrator/prompt.rs, and integrations_agent/prompt.rs to inject a "Connected accounts:" block. Add it to any new agent prompt that needs Composio context.

Agent Timeout & Cancellation (Issue #715)

• Frontend silence timer, not a wall-clock limit — armSilenceTimer in app/src/pages/Conversations.tsx fires if 120s (fixed to 600s) pass with zero inference progress events. It re-arms on every tool_call, tool_result, iteration_start, etc., so long-running tool chains that keep emitting events are not cut off.
• Rust-side HTTP timeout is separate — src/openhuman/providers/compatible.rs sets a 120s reqwest client timeout on LLM calls. Not changed in #715; relevant if a single LLM round-trip itself stalls for >2 min.
• Manual cancel path — chatCancel() in app/src/services/chatService.ts → openhuman.channel_web_cancel RPC → cancel_chat() in src/openhuman/channels/providers/web.rs. Fully implemented; the silence timer is an automatic fallback.

Webhook & Cron Triggers (Issue #726)

• Webhook bus was hardcoded 410 — src/openhuman/webhooks/bus.rs WebhookRequestSubscriber::handle() returned 410 "skill runtime removed" for ALL incoming webhooks. Now routes to echo/agent/skill/404 based on TunnelRegistration.target_kind.
• WebhookRouter access from bus.rs — Router lives in SocketManager::shared.webhook_router (was pub(super)). Added pub fn webhook_router(&self) accessor on SocketManager; bus.rs reaches it via global_socket_manager().webhook_router().
• TriggerSource enum: three update points — Adding new variants requires updating: (a) slug() match in envelope.rs, (b) exhaustive test match, (c) handle_triage_evaluate string match in agent/schemas.rs (uses p.source.as_str(), not the enum directly).
• CronJobTriggered/CronJobCompleted were never published — Defined in events.rs and used in tests but never emitted. Now published by execute_and_persist_job() in scheduler.rs. Adding fields to these variants requires updating ~5 construction sites: cron/bus.rs, composio/bus.rs, tree_summarizer/bus.rs, channels/proactive.rs, and events.rs tests.
• Webhook ops were all stubs — list_registrations, list_logs, clear_logs, register_echo, unregister_echo in ops.rs all returned empty. Now backed by the real router via a get_router() helper.
• GGML_NATIVE=OFF for cargo check — Sidestepping the whisper-rs macOS Tahoe build blocker for cargo check: GGML_NATIVE=OFF cargo check --manifest-path Cargo.toml. Allows compilation checks without the cmake failure.

Agent Runtime Behavior

• sandbox_mode = "read_only" in agent.toml is metadata only — Never enforced at runtime. Actual security policy comes from config.autonomy (global), defaulting to Supervised. Adding write tools to a read-only agent works at runtime but violates documented intent.
• max_iterations hard-fails, not graceful truncation — When the welcome agent (or any agent) hits max_iterations, tool_loop.rs:705 calls anyhow::bail!. There is no graceful truncation. Budget iterations carefully.
• Archivist agent auto-extracts memory — It processes conversation history and persists preferences/facts into user_profile automatically. Agents do not need to explicitly call memory_store to persist conversational insights.
• cargo check / cargo test fails on main (llama.cpp cmake) — llama.cpp's cmake build script uses -mcpu=native, which is unsupported on Apple clang 21+ with --target=arm64-apple-macosx. Pre-existing issue on main, not branch-specific. Frontend checks (typecheck, lint, format) are unaffected. Workaround: set GGML_NATIVE=OFF (same fix as whisper-rs above).

Cron Scheduler

• Cron loop was never spawned — tokio::spawn(cron::scheduler::run(config)) was missing from src/core/jsonrpc.rs. Added after the update scheduler spawn, gated on config.cron.enabled. Without it, scheduled jobs never auto-fire at startup (issue #830).

Build & Tooling Gotchas

• pnpm typecheck script was renamed — Check app/package.json for the current name; as of issue #830 work, use pnpm workspace openhuman-app compile for tsc checks.
• PR #745 (command palette) merged without its deps — @radix-ui/react-dialog, cmdk, and @testing-library/user-event are missing from package.json. Install them if tsc fails after syncing main.
• Pre-push hooks fail on upstream lint warnings — ESLint warns on setState in effects and unused eslint-disable directives inherited from upstream. Use --no-verify only when the lint errors are pre-existing upstream issues, not new code.
• pnpm tauri icon <source.png> generates all platform icons at once — Produces .icns, .ico, all PNG sizes, Windows Store tiles, and iOS/Android sets. Use this instead of manual sips/ImageMagick resizing.
• tauri-cef submodule update can fix missing Tauri runtime modules — e.g. updating to f75bc21f5 added the missing tauri_runtime_cef::audio module that was causing pre-push hook compile failures on the Tauri shell. When the shell fails to compile with a missing module error, check if the submodule needs updating.
• git add must run from repo root — Staging paths like app/public/... with git add from inside app/ won't match. Always run git add from /Users/megamind/tinyhuman/openhuman-claude.
• Brand kit assets live at app/public/brand/ — Copied there during session work; original source is in ~/Downloads/Brand kit/. Not auto-synced; re-copy manually if Downloads content changes.
• pnpm test:coverage ENOENT on coverage/.tmp/coverage-0.json — Race condition in coverage file collection; flaky, not reproducible every run. Use pnpm debug unit instead — runs Vitest without coverage, faster and reliable for iteration.

Mascot Native Window (macOS)

• Not a Tauri window — The floating mascot is a native NSPanel + WKWebView in app/src-tauri/src/mascot_native_window.rs. It uses ignoresMouseEvents=true (click-through); interaction is detected by polling NSEvent via a Foundation timer. macOS-only, uses objc2 bindings.
• MainThreadOnly import must stay — Required by WKWebView::alloc() and other AppKit allocators even if not explicitly referenced in user code. Removing it causes compile errors.
• NSEvent::pressedMouseButtons not in typed objc2-appkit bindings — Must be called via msg_send!(objc2::class!(NSEvent), pressedMouseButtons) instead of the typed API.
• WKWebView IPC via evaluateJavaScript — The mascot webview is NOT a Tauri runtime; Tauri invoke/emit do NOT work. Rust-to-mascot communication uses msg_send!(webview, evaluateJavaScript:completionHandler:) to dispatch new CustomEvent(...) on window. React listens with addEventListener. This is NOT subject to the CEF JS injection ban (that only applies to webview_accounts/ third-party origins).
• MascotCharacter sleeping prop — Drives the sleep animation (eye close + Zzz). sleepStartSec and sleepFullSec are hardcoded at 2.5s and 4.0s — they are NOT configurable props. Only toggle sleeping: boolean.
• FACE_PRESETS is a strict Record — Typed as Record<Exclude<MascotFace, 'normal'>, FacePreset> in Ghosty.tsx. Adding a new MascotFace union variant requires adding a matching entry to FACE_PRESETS or it won't compile.
• _webview in spawn_drag_timer — The WKWebView captured in the drag timer closure was originally unused (prefixed _). It can be used for evaluateJavaScript calls during the hover polling loop (e.g. to trigger blink/wake events from Rust).
• FrameProvider loops — sleep animation resets — FrameProvider uses frame % durationInFrames so animations loop. Default DURATION_FRAMES = FPS * 6 (6s). Sleep animation completes at 4s, then eyes re-open at 6s when frame resets to 0. Fix: use a much longer durationInFrames for sleep face (e.g. FPS * 600) so the loop never triggers while sleeping.
• Hover detection needs circular hitbox — The mascot panel is 79x79 but the character is visually circular. Using the full AABB (cursor_in_panel) for hover triggers false positives when cursor is in a panel corner. Use distance-from-center check instead. Also suppress hover events for ~1s after panel shows to let the webview load.

Google Analytics (Issue #1479)

• react-ga4 injects a <script> tag at runtime — It appends a gtag.js <script> to <head> dynamically. This works because tauri.conf.json CSP has https: in default-src and connect-src. If CEF ever tightens script-src separately, switch to GA4 Measurement Protocol (pure HTTP POST, no script injection).
• Analytics module pattern — app/src/services/analytics.ts is the single owner of initGA, trackPageView, trackEvent, plus an ALLOWED_EVENTS allowlist. Never call ReactGA directly from components; go through this module.
• Triple gate before any GA call — isAnalyticsEnabled() (user consent) AND GA_MEASUREMENT_ID env var present AND !IS_DEV. All three must pass or tracking is silently skipped.
• Route tracking location — useLocation() effect wired in AppShell (not individual pages). All page views emit from one place.
• Capability catalog must stay in sync — src/openhuman/about_app/catalog.rs needs an entry when a new user-visible feature ships. GA was added there as part of issue #1479.

PR Checklist CI

• N/A items need a checked checkbox — scripts/check-pr-checklist.mjs requires - [x] N/A: <reason>. Using - [ ] N/A: (unchecked) fails the check even though the text starts with "N/A:".

Config System (Rust)

• Config corruption recovery — parse_config_with_recovery in src/openhuman/config/schema/load.rs: try primary → try .bak → archive corrupt file → Config::default(). Guarantees the app always starts even with a corrupt config.
• New config fields must use #[serde(default = "fn_name")] — Bare #[serde(default)] gives 0/false, not the meaningful domain default. Define a named fn returning the correct value and reference it by name.
• .bak is now permanent — Config::save() no longer deletes .bak on success. It always reflects the last-known-good config before the most recent write.
• load_from_default_paths has zero callers — Debug utility only; not user-facing.
• Config test module path — openhuman::config::schema::load::tests. Run with cargo test -- config::schema::load::tests.

Environment

• Core port — 7788 (default; in-process inside Tauri host). Check with lsof -i :7788.
• pnpm core:stage — no-op (sidecar removed in PR #1061). Use pnpm dev:app for full Tauri+core dev.
• Kill stuck processes — lsof -i :7788 then kill <PID>. Useful when dev:app reports a stale listener and you want to force a fresh boot rather than relying on the handle's auto-recovery.
• Skills runtime rebuilt (PR #2707) — QuickJS is gone, but skills now run as orchestrator-focused agents via skills_run RPC. Default skills live in src/openhuman/skills/defaults/<id>/ with skill.toml + SKILL.md, registered in registry.rs DEFAULT_SKILLS const. Seeded into <workspace>/skills/ on boot (idempotent, non-destructive). Bundled defaults: github-issue-crusher, dev-workflow. Skills run with 200 iteration cap and full web access.
• Codegraph tools (PR #2707) — codegraph_index and codegraph_search registered in src/openhuman/tools/ops.rs. Implementation in src/openhuman/codegraph/ — tree-sitter extraction, SQLite FTS5, dense embeddings, RRF fusion. Auto-indexes on first search.
• Tool names are exact — Always check src/openhuman/tools/ops.rs for authoritative names. Key ones: edit (not edit_file), composio (not composio_execute), codegraph_index, codegraph_search.
• cron_add RPC — Was missing from schemas.rs (only existed as agent tool). Now exposed as openhuman.cron_add. Frontend wrapper: openhumanCronAdd() in app/src/utils/tauriCommands/cron.ts.
• Worktree pnpm build rolldown fix — Worktrees can miss @rolldown/binding-darwin-arm64. Fix: pnpm install --force.

Artifacts Domain (Issue #2776)

• Filesystem-backed persistence, no SQLite — src/openhuman/artifacts/ stores JSON metadata (meta.json) + binary blobs under <workspace_dir>/artifacts/<uuid>/. Pattern mirrors memory/ops/files.rs but simpler.
• "ai" namespace in controller registry — RPC methods are openhuman.ai_list_artifacts, openhuman.ai_get_artifact, openhuman.ai_delete_artifact. Future ai_* methods should use this same namespace.
• Two-layer path validation required — (1) validate_artifact_id rejects empty strings, /, \, .., absolute Unix paths, Windows C: and UNC \\ paths; (2) assert_within_root canonicalizes and checks containment. Replicate this pattern for any new filesystem-backed domain.
• cargo test --lib required for lib crate tests — cargo test -p openhuman -- "artifacts" lists tests but filters to 0. Must use cargo test -p openhuman --lib -- "artifacts" because tests are in the lib crate, not integration test binaries.

Rust Testing Patterns

• Memory tree tests filter — cargo test -p openhuman -- "memory::tree" runs the memory tree unit tests (602 tests); full module paths are openhuman::memory::tree::ingest::tests::* and openhuman::memory::tree::canonicalize::email_clean::tests::*.
• cargo fmt --all — Required after codecrusher generates Rust; it doesn't always produce perfectly formatted output and CI will reject unformatted code.
• PR quality scripts are soft checks — scripts/check-pr-checklist.mjs and scripts/check-coverage-matrix.mjs exit cleanly with summary lines; CI treats them as advisory, not blocking.
• ceil_char_boundary — Safe string slicing utility at src/openhuman/util.rs; use this throughout the codebase instead of raw byte-index slicing to avoid UTF-8 panics.
• Global static cache tests need a reset guard — When testing code that reads/writes a Lazy<Mutex<Option<...>>> global cache, use a struct CacheResetGuard; impl Drop for CacheResetGuard { fn drop(&mut self) { *CACHE.lock() = None; } } pattern so each test starts clean. See SnapshotCacheResetGuard / CacheResetGuard in ops_tests.rs.
• Test assertions must match the actual dummy value — When a builder (e.g. build_dummy_runtime_snapshot()) wraps degraded_runtime_snapshot(), assert against dummy.field rather than a hardcoded string (e.g. "idle" vs the actual "degraded") to verify round-trip correctness without false mismatches.
• composio::action_tool::tests::mode_toggle_between_calls_is_observed is flaky in full suite — Fails intermittently due to shared global composio session state; passes in isolation. Pre-existing; not caused by snapshot perf work.
• GLOBAL_MEMORY_TEST_LOCK only serializes test bodies, not background workers — Background ingestion spawned by a prior test can still be running when the next test acquires the lock. Call state.reset_for_test() at test start (after acquiring the lock) to clear accumulated queue_depth/running state; do not rely on delta assertions alone.
• IngestionState::reset_for_test() is #[cfg(test)]-gated — Lives in src/openhuman/memory/ingestion/state.rs. Zeroes queue_depth (AtomicUsize) and clears running/current fields in the snapshot while preserving completion history. This is the canonical reset for any test asserting exact queue or running state.
• cargo-llvm-cov widens SQLITE_BUSY window — Flakes that only appear under coverage (cargo-llvm-cov) but not plain cargo test are usually (a) a SQLite connection missing busy_timeout, or (b) shared global state not reset between tests. Always set busy_timeout on new SQLite connections (see pattern below).
• All new SQLite connections must set busy_timeout = 15s — Call conn.busy_timeout(Duration::from_secs(15)) immediately after Connection::open(), before any execute_batch(). Pattern set by chunks/store.rs (SQLITE_BUSY_TIMEOUT) and now also used by memory_store/unified/init.rs (fixed in issue #2722). Without it, concurrent ingestion + test writes produce SQLITE_BUSY under cargo-llvm-cov.

App State Snapshot (Issue #2155 — first-launch perf)

• build_runtime_snapshot was serial, now parallel — The four subsystems (screen intelligence, local AI, autocomplete, service status) in src/openhuman/app_state/ops.rs ran sequentially. Fixed with tokio::join!. Also added a 2s TTL cache (RUNTIME_SNAPSHOT_CACHE) so repeated polls within the TTL skip recomputation.
• service::status is sync — must use spawn_blocking — crate::openhuman::service::status(config) may shell out to launchctl. Wrap it in tokio::task::spawn_blocking when called from an async context.
• autocomplete::global_engine().status() calls Config::load_or_init() internally — Avoid this inside snapshot code. Use the new status_with_config(config) method which accepts an already-loaded config.
• Per-stage snapshot timeouts — AUTH_FETCH_TIMEOUT = 5s and RUNTIME_SNAPSHOT_TIMEOUT = 10s are constants in ops.rs; they sum to 15s, well under the 30s frontend RPC timeout.

Project Board & Issue Queries

• Project #2 paginates at 100 items — Board has 627+ items. Use GraphQL cursor pagination to find all open P0 issues; a single query only returns the first 100.
• jq regex \s+ causes parse errors — Use plain test("#NNNN") to check if a PR/issue body references an issue number. \s+ in jq regex triggers parse errors.
• Most open P0s are security or Linux AppImage GLIBC issues — When triaging P0s, filter for those categories first.
• Project #2 shows only closed items on the board view — Use gh issue list --repo tinyhumansai/openhuman --state open --assignee "" to find unassigned open issues instead of querying the project board.
• Check linked PRs via timeline API, not body regex — gh api repos/tinyhumansai/openhuman/issues/$N/timeline --paginate | jq '[.[] | select(.event == "cross-referenced" and .source.issue.state == "open")] | length' is more reliable than searching issue body text for PR references.

Git Submodules

• tauri-cef and tauri-plugin-notification are git submodules — When upstream/main updates them, fix with git submodule update --remote --checkout, not by manually patching the vendored crate.

Pre-existing Test Failures

• composio::action_tool::tests::factory_routes_through_direct_when_mode_is_direct fails in cargo test -p openhuman — Pre-existing failure unrelated to WhatsApp or any recent branch work. Do not attempt to fix unless explicitly tasked. Also intermittently flaky when run as part of the full suite — see "Pre-existing Flaky Tests" section.

Workflow Gate (must not skip)

• Steps 4–6 of workflow/00-full-workflow.md are mandatory before committing — Step 4: architectobot verify. Step 5: full checks (pnpm test:coverage, pnpm build, bash scripts/install.sh --dry-run, PR quality scripts). Step 6: memory-keeper. Skipping any of these violates the workflow contract.
• Encode architectobot answers in the codecrusher prompt — When the architectobot plan includes clarifying questions and the user approves specific answers, embed those decisions as explicit constraints in the codecrusher prompt so the agent doesn't re-ask.

Security Policy

• Path validation entry point — src/openhuman/security/policy.rs exposes validate_path / validate_parent_path. All file I/O path validation must go through this API. is_path_string_allowed() is a string-only first pass, not sufficient on its own.
• validate_parent_path before create_dir_all — For write operations, validate_parent_path MUST be called before any create_dir_all call. Calling it after allows symlink attacks to create directories outside the workspace before the security check fires (Issue #1927).
• Tool callers must use validate_path / validate_parent_path — All tool implementations under src/openhuman/tools/impl/filesystem/ must use these functions, not the legacy is_path_allowed / is_resolved_path_allowed.
• Security policy test filter — Run only security policy tests with: cargo test -p openhuman -- "security::policy". Runs the 100 tests in src/openhuman/security/policy_tests.rs cleanly.

Pre-existing Flaky Tests

• composio::action_tool and agent::harness::session::turn intermittent failures — These tests fail randomly when run as part of the full suite (likely shared state or timing), but pass individually. Not related to security/policy changes. Do not treat as blockers for security-module PRs.

Windows OAuth Deep Link (Issue #2562)

• Three-layer fix: (1) named-pipe IPC in deep_link_ipc_windows.rs — secondary process forwards openhuman:// URL to primary via \\.\pipe\com.openhuman.app-deeplink, 40 retries × 50ms; (2) loopback OAuth server in loopback_oauth.rs — RFC 8252 one-shot 127.0.0.1:53824, preferred path that eliminates deep link dispatch entirely; (3) Linux analog in deep_link_ipc.rs — Unix domain socket at $XDG_RUNTIME_DIR/com.openhuman.app-deeplink.sock.
• OAuthProviderButton.tsx loopback flow — tries loopback first, sets redirectUri for backend, awaits callback, rewrites http://127.0.0.1:PORT/auth?... → openhuman://auth?... → handleDeepLinkUrls. Falls back to deep link if bind fails.
• Pipe binding location — primary binds the named pipe in lib.rs right after the mutex guard (grep bind_and_listen); drain_pending_urls() wired in setup() (grep it — line numbers drift).
• Issue was already fixed before we picked it up — PRs #2469, #2511, #2550 had already merged the fix. Our contribution was extracting classify_request as a pure function and adding 11 Rust unit tests.
• Pure-function extraction pattern — when async/AppHandle-gated Tauri code is untestable, extract a classify_request(head, expected_state, bound_port) -> RequestOutcome pure function returning an enum. Enables comprehensive unit tests with zero Tauri context. RequestOutcome has 4 variants: AuthCallback, StateMismatch, NotFound, MethodNotAllowed.

Port Conflict Recovery (Issue #2617)

• Port fallback already in pick_listen_port — src/openhuman/connectivity/rpc.rs tries ports 7789–7798 when 7788 is busy. Gap was: frontend getCoreRpcUrl() cached the URL on first resolution so it never picked up the fallback port, and stale-process reaping was macOS-only.
• process_recovery.rs is platform-gated — reap_stale_openhuman_processes had only a macOS impl. Linux uses /proc/<pid>/cmdline; Windows uses wmic process get. Tests for each platform's parsing logic live in the same file, following the existing macOS test pattern.
• recover_port_conflict is a Tauri IPC command, not JSON-RPC — Rust E2E test for port fallback lives in tests/json_rpc_e2e.rs and calls pick_listen_port directly: bind port 7788 with a std::net::TcpListener (std, not tokio) to simulate conflict, confirm fallback, then serve via tokio::net::TcpListener::from_std(pick_result.listener.into_std()).
• BootCheckTransport is the right hook for frontend recovery — app/src/lib/bootCheck/index.ts is the injection point for new recovery capabilities; don't add them directly to the BootCheck component.
• i18n locales are single flat files — Each locale is one file at app/src/lib/i18n/<locale>.ts (en.ts is the source of truth; the old chunks/<locale>-N.ts layout was retired). New keys must be added to all 13 locale files simultaneously; pnpm i18n:check enforces key parity.
• Workflow folder — workflow/ at repo root has 5 markdown files (00–05) defining the full PR workflow: pick issue → architectobot plan → user approval → codecrusher → architectobot verify → checks → memory-keeper → commit → push/PR.

Channel Event Workspace Routing (Issue #2602)

• Workspace identity is PathBuf — Represented as the workspace directory path on ChannelRuntimeContext as ctx.workspace_dir: Arc<PathBuf>. Use ctx.workspace_dir.as_ref().clone() at publish sites. There is no abstract WorkspaceId type.
• DomainEvent workspace routing contract — Publisher populates workspace field from context; subscriber compares against self.workspace_dir and early-returns with log::debug! on mismatch. Follow this pattern for any workspace-scoped DomainEvent variant.
• ChannelMessageReceived and ChannelMessageProcessed carry workspace_dir — Added in PR for issue #2602. Guards in ConversationPersistenceSubscriber (memory_conversations/bus.rs) and TelegramRemoteSubscriber (telegram/bus.rs) prevent cross-workspace persistence during login/workspace-change races.

Pre-existing Upstream Failures (from issue #2602 session)

• Upstream main has 5 Vitest failures and 4 TypeScript compile errors — Caused by missing iOS experimental dependencies: @noble/ciphers/chacha, @noble/ciphers/webcrypto, qrcode.react, @tauri-apps/plugin-barcode-scanner. Breaks pnpm compile, pnpm build, pnpm test:coverage on a clean checkout. Always verify by stashing changes and running checks on the base branch before blaming your PR.
• cargo fmt must run after codecrusher — codecrusher does not reliably produce cargo fmt-clean Rust. Always run cargo fmt --manifest-path Cargo.toml after codecrusher finishes and before committing.

TaskKanbanBoard (Issue #3347 — frontend-only subset)

• 5 columns, 7 statuses — COLUMN_DEFS in TaskKanbanBoard.tsx defines columns todo / awaiting_approval / in_progress / blocked / done. columnFor() maps the two virtual statuses: ready → in_progress (renders a "Ready to start" sage badge) and rejected → done (renders a "Rejected" coral badge). Anyone adding a new TaskBoardCardStatus must update both COLUMN_DEFS and columnFor().
• Native HTML5 drag-and-drop only — No DnD library (@dnd-kit, react-beautiful-dnd). Cards set dataTransfer key application/x-task-card-id; columns handle onDragOver / onDragLeave / onDrop. Arrow buttons are the touch/accessibility fallback. Do not add a DnD library.
• mutatingCardId is optional on TaskKanbanBoard — IntelligenceTasksTab sets/clears it in finally blocks around personal-board mutations; Conversations.tsx (second consumer) does not pass it. Both consumers must keep new props optional to avoid compile failures.
• Bare vitest run <path> fails with "document is not defined" — Must use --config test/vitest.config.ts (or pnpm debug unit <relative-path>) to load the jsdom environment. Piping pnpm exec vitest run through | tail also buffers all output until completion, hiding progress — use the debug runner instead.

Memory Sync Sources — Defaults ON + Per-Source UI (Issue #3293)

• Conservative caps registry — composio_defaults_for_toolkit(toolkit) -> (max_items, sync_depth_days) in src/openhuman/memory_sources/registry.rs is the single source of truth. Values: gmail 100/30, slack 50/14, notion 30/30, linear 50/30, clickup 50/30, github 50/30, generic 30/14. Non-Composio defaults (GithubRepo 10PR/10issue/50commit, RSS 20, Twitter 7d) live in apply_kind_defaults in rpc.rs.
• upsert_composio_source now defaults ON — Registers enabled: true with caps applied from the registry. Previously registered enabled: false with no caps.
• Cap enforcement: 3 construction sites, not 1 — ProviderContext (memory_sync/composio/providers/types.rs) carries max_items/sync_depth_days. All three sites must populate caps from the registry entry: composio/mod.rs run_connection_sync, composio/periodic.rs, composio/bus.rs. Each provider reads ctx.max_items/ctx.sync_depth_days for pagination + date clamping. Shared helpers: pages_for_max_items / epoch_floor_from_depth in providers/helpers.rs.
• First-sync gotcha on new connections — bus.rs on_connection_created fires BEFORE upsert_composio_source, so caps are (None, None) on the brand-new connection's first sync — bounded only by internal MAX_PAGES, not registry caps. Documented in code; intentional.
• memory_sources_apply_all_in RPC — Zero params → { sources, sync_triggered }. Enables all sources, clears caps to None (falls back to internal MAX_PAGES ceilings, ~500 items, not truly unlimited), triggers sync per source.
• Retroactive migration — apply_composio_source_caps_migration in reconcile.rs, guarded by Config.composio_source_caps_migrated: bool (#[serde(default)]), runs once from ensure_composio_sources. Only touches Composio entries with !enabled && max_items.is_none() && sync_depth_days.is_none() — never overwrites user-customized caps.
• MemorySourcePatch was missing limit fields — max_commits, max_issues, max_prs were absent from MemorySourcePatch, update_source, and the update schema. Also missing from the TS MemorySourceEntry interface in memorySourcesService.ts. Both were fixed in this issue.
• Per-source settings UI — SourceSettingsPanel.tsx (sibling of MemorySourcesRegistry.tsx) with a KIND_FIELDS map driving which limit fields appear per source kind. Empty input = omit from patch (use default); number = set. "All In" button uses ConfirmationModal (primary-500 prominent). Toasts via existing onToast prop.
• pnpm i18n:english:check is pre-failing — Exits with total unexpected English: 1312 on a clean base tree. Confirm your new keys aren't among the failures rather than expecting exit 0. pnpm i18n:check (key parity) is the real gate.

Memory Tree Sync & Raw-Archive Reconcile (branch fix/memory-tree-sync-reconcile)

• Tree scope ≠ raw-archive id for GitHub — github:owner/repo (tree registry key) and github.com/owner/repo (raw archive id) slugify to DIFFERENT raw/ dirs (github-owner-repo vs github-com-owner-repo). Any code resolving a source's raw dir must use memory_sources::sync::SourceScope { tree_scope, archive_source_id } — deriving from the tree scope scans the wrong directory and silently reports "nothing to do".
• Raw-file coverage gate — mem_tree_ingested_sources with source_kind = "raw_file", source_id = <rel path with .md> (memory_store/chunks/store.rs: mark_raw_paths_ingested / filter_raw_paths_not_ingested). Written by github sync + rebuild after each summary batch lands. raw_coverage() in memory_sync/sources/rebuild.rs backfills the gate once per scope from L1 summary child labels (commit:<sha> / issue:<n> / pr:<n> / file stems) before diffing disk vs gate.
• Workspace periodic scheduler — memory_sync/workspace/periodic.rs (started in core/jsonrpc.rs next to the Composio one) is the ONLY thing auto-syncing github_repo/folder/rss_feed/web_page sources; the Composio loop walks Composio connections exclusively. Cadence: config.memory_sync_interval_secs (Some(0) = manual only, default 24h), due-check via in-memory fired-at map + persisted sync-audit fallback.
• Queue self-heal retries transient failures only — memory_queue/scheduler.rs calls store::requeue_transient_failed every 3h (excludes failure_class = 'unrecoverable' so bad keys/budget don't retry-loop). Full reset stays manual via memory_tree_retry_failed RPC.
• openhuman.memory_sources_reconcile RPC — Reports per-scope raw coverage {total_raw_files, covered, pending}; execute: true spawns background incremental rebuild. Same reconcile auto-runs after every sync via check_and_rebuild_tree.
• L0 seal gate is token-only by design — should_seal has NO sibling-count fallback at L0 (docs used to lie about this); small buffers seal via the 3-hourly flush_stale path. L≥1 gates on SUMMARY_FANOUT (10) — an L1 buffer holding <10 summaries is healthy, not stuck.

Intelligence page scroll ownership (#4267)

• app/src/pages/Intelligence.tsx is rendered ONLY by Brain.tsx (/settings/intelligence redirects to /brain?tab=intelligence), inside a card with h-full overflow-hidden (Brain.tsx ~line 255). A code comment there asserts the embedded panels "own their own scroll directly" — true for the sibling PanelPage/PanelScaffold panels (MemoryDataPanel etc.), but Intelligence previously returned a bare <div className="z-10 relative"> with NO height/scroll, so its content overflowed the card and was clipped with no scrollbar.
• Fix pattern (matches PanelPage): root flex h-full min-h-0 flex-col → pinned SettingsHeader wrapper (flex-shrink-0) → body min-h-0 flex-1 overflow-y-auto wrapping ChipTabs + the content card. ToastContainer/ConfirmationModal stay as siblings OUTSIDE the scroll region. The scroll body carries data-testid="intelligence-scroll" for the regression test.
• Gotcha for any panel embedded in Brain's Knowledge/Memory card: it must own its own vertical scroll (the card is overflow-hidden). Don't rely on a parent scroll — there isn't one.