Ship: BB SDK (core + realtime), custom models, app scaffold, CC workflows + adversarial-review fixes#64
Merged
Merged
Conversation
…+ CLI consumer New @bb/sdk package: typed per-area SDK (threads/apps/hosts/environments/ projects/providers/managers/status/guide/replay) with a transport abstraction that runs in node (CLI), browser, and injected app runtimes. Types derive from @bb/server-contract / @bb/domain. CLI refactored to a thin consumer. window.bb is regenerated from the real createInjectedBbSdk (single source of truth, no hand-rolled string), exposing the full CLI-power surface with object- arg data/message APIs and websocket-backed data.onChange realtime. No back-compat shims (prototype). Phase 2 (bb.on realtime event map) deferred with seams left. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
The thread/manager header sits in an Electron title-bar drag region, so the "..." ThreadActionsMenu trigger was swallowed as a window drag. Apply the existing MACOS_WINDOW_NO_DRAG_CLASS to the header trigger only (sidebar usage unchanged). Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
The native WebContentsView bounds were driven only by an async renderer
pipeline (ResizeObserver/window resize -> rAF -> IPC -> setBounds), so during
a synchronous OS window resize the view trailed the window edge by >=1 frame +
an IPC hop. Electron 41's WebContentsView has no setAutoResize.
Introduce a resize-invariant layout descriptor ({left, top, rightInset,
bottomInset}) the renderer emits only on layout-shape changes. The main process
caches it per view and, on the host BrowserWindow will-resize/resize events,
re-projects bounds from getContentBounds() and calls setBounds() synchronously
in lockstep with the OS resize. Preserves the prior keep-alive (#94),
panel-resize (#96), and containment (#128) behavior.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Delete manager templates end to end — server route + storage-template service, host-daemon command handler, the domain/server/host-daemon contract types and fields (templateName / managerTemplateName / ManagerTemplateName / GET /manager-templates / host.list_manager_templates), the bb manager CLI options, the frontend ManagerTemplatePicker + compose-view usage, the bb-guide-manager-templates chapter (regenerated), and all related tests. Managers are still created normally; they now start from default/empty thread storage with no template selection or seeding. Existing on-disk ~/.bb/manager-templates/ user data is left intact and is now unused. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
…degen Replace the static single-file app scaffold with a full Vite + React + TypeScript Todo app that showcases the window.bb SDK and live data-binding: - editable source/ + pre-built public/ (served web root, flat refs), copied into each new app by the server (excludes dev-only screenshots/report dirs); - an add-todos skill + README; - SDK types are GENERATED from @bb/sdk (self-contained ambient window.bb d.ts, no imports) via packages/sdk/scripts/generate-app-globals-dts.mjs, with a drift-guard test so the vendored template types can't silently diverge. Polished UI pass (lucide, bb oklch tokens, responsive); Live-pill + bound-data showcase removed; error notices retained. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
electron-builder's default node_modules .d.ts pruning stripped the app-scaffold-template's generated source/src/bb-sdk.d.ts from app.asar, so apps created in the packaged desktop app scaffolded without their window.bb types. Add a dedicated files FileSet for the template tree so it ships verbatim, without relaxing .d.ts pruning elsewhere. Regression test added. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Add a typed realtime subscription API to @bb/sdk: bb.on({event, ...scope,
callback}) returning an idempotent unsubscribe, across thread/project/
environment/host/system (config/apps) / app / app-data (changed+resync) plus a
realtime:connection lifecycle event. One websocket per SDK instance with
ref-counted subscriptions, reconnect + resubscribe, and contract-derived
payloads (no SDK-only mirrors). window.bb.data.onChange now rides the shared
socket with full Phase 1 parity (subscribe-before-replay, buffer, version
dedupe, resync + reconnect replay). Server broadcasts app:changed alongside
system:apps-changed; app-data stays per-application.
Hardened after dual review: reject orphaned socket-ready promise on
close-before-open; connection listeners are observers (no socket ownership);
reset reconnect intent on idle close (no double-replay); reconnect replays
before emitting connected; fail-safe outgoing broadcast validation.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Local electron-builder runs previously forced identity=null and CSC_IDENTITY_AUTO_DISCOVERY=false unless the full CI secret set was present, so every locally packaged bb.app shipped the prebuilt Electron's invalidated adhoc linker signature (spctl: "code has no resources but signature indicates they must be present"). macOS provenance-tracks such bundles, which forces syspolicyd to evaluate and journal every exec in the app's process tree — observed pegging syspolicyd at ~380% CPU and stalling process launches system-wide. Signing now resolves to one of three modes: - environment: full CI secrets — sign with the provided cert + notarize (published-release path, unchanged) - keychain: no secrets — sign via keychain auto-discovery, skip notarization (locally built apps never carry the quarantine xattr, so notarization is unnecessary; a valid Developer ID seal is what lets Gatekeeper cache assessments and keeps fresh launches out of the provenance sandbox) - disabled: no secrets and CSC_IDENTITY_AUTO_DISCOVERY=false — explicit unsigned build (CI workflow-artifact-only path, unchanged) Validated: packaged app deep-verifies strict with sealed resources (6105 files), spawn-helper and better_sqlite3.node carry the Team ID, and the hardened-runtime app fully boots (bridge/server/daemon all re-exec the signed binary via ELECTRON_RUN_AS_NODE). Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Add a customModels key to the bb-app managed config so users can register
model ids the provider catalog doesn't offer (e.g. non-public preview
models):
{ "customModels": [{ "providerId": "claude-code", "model": "<id>",
"displayName": "<label>" }] }
The server appends them to the daemon-reported model list in
/system/execution-options — including when the provider catalog fails to
load — with catalog metadata winning on model-id collision (selected-only
entries are promoted rather than shadowed). Reasoning ladders are
per-provider: claude-code gets the full low-max ladder, while codex and pi
cap at xhigh since both reject "max" provider-wide. providerId is
validated by agentProviderIdSchema at the schema boundary so the launcher
and server agree on validity and reload 422s name the offending field.
The launcher preserves customModels across managed-config writes
(pruneManagedConfig previously dropped unknown keys) and surfaces entries
in `bb-app config list`. Reasoning-effort constants move from
@bb/agent-runtime to @bb/domain so the server can build AvailableModel
entries without depending on the runtime package.
No frontend changes: config-changed notifications already invalidate the
execution-options queries, so open pickers refresh on reload.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Translate the agent SDK's task event family (task_started/task_updated/
task_progress/task_notification) into a new backgroundTask thread item:
dynamic workflows render as a live timeline row with phase groups and
per-agent progress, folded from workflow_progress delta batches and
throttled adapter-side. Progress/completion events are thread-scoped so
tasks that outlive their spawning turn stay pagination-safe; timeline
windows backfill the latest task state for in-window items.
Lifecycle ownership: the adapter settles open tasks on thread/resume and
provider process exit; the server settles dangling items when a daemon
session re-registers with a new instance id; superseded progress rows are
pruned keep-latest-while-pending.
Adds the "ultracode" reasoning level (ranked between xhigh and max,
reconciling down to xhigh on model switch) for xhigh-capable claude-code
models. The server-owned workflowsEnabled policy flows explicitly through
the daemon contract (protocol v31); the adapter decomposes ultracode into
effort "xhigh" plus flag-tier settings {enableWorkflows, ultracode}.
Also: status:null now completes dangling contextCompaction items, and
session_state_changed/tool_progress/tool_use_summary are classified so
they no longer surface as provider/unhandled debug rows. Fixtures are
captured from a real workflow run driven through the agent SDK.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Fixes from a 99-agent adversarial review of e0b03c6d7 (41 raw findings → 28 confirmed after 3-lens verification): Client lifecycle (packages/sdk/src/realtime-client.ts): - connectSocket constructs the socket before creating the socket-ready promise, so a sync throw (bad URL/factory) can no longer orphan a pending promise into an unhandled rejection; the reconnect timer body also contains sync throws instead of crashing the process - onclose always records reconnect intent and emits 'disconnected' even while a stale backoff timer is pending, and a listener-driven connect cancels the pending timer — previously a close during backoff silently skipped the reconnect replay (lost app-data events, Phase 1 parity regression) and the orphaned timer escalated the delay while connected - idle close that cancels a pending reconnect now emits a terminal disconnected event so observers don't wait forever on a promised retry - dispatch iterates a listener snapshot: listeners added inside a callback no longer receive the in-flight event (preserves subscribe-then-replay ordering); replay delivery loops re-check listener.active so unsubscribe stops deliveries immediately - buffered-event flush uses a per-path last-version rule: when the replay snapshot already delivered a path's final buffered state, all buffered events for that path are skipped — the old exact-match dedupe could re-deliver an older value last, leaving consumers permanently stale (versions are content hashes, so ordering cannot disambiguate) - reconnect now notifies app-data:resync subscribers (broadcasts may have been missed) before replaying and emitting reconnected - late realtime:connection observers receive the current state as a microtask snapshot instead of observing nothing until the next transition Version-skew tolerance (domain/server-contract): - changed-message and ThreadChangeMetadata types are now z.infer-derived from their schemas — schema/type drift is a compile error instead of a silent fail-closed realtime blackout at the hub's outgoing gate - new lenient inbound schemas (changedMessageLenientSchema, serverMessageLenientSchema): clients strip unknown fields and filter unknown change kinds instead of dropping whole messages when talking to a newer server; SDK and web app now parse inbound traffic with them (the strict schemas keep guarding the server's outgoing boundary) - changes arrays are readonly in the message types; entity events are delivered as shared objects with mutation blocked at compile time - dropped the dead optional id from AppChangedMessage (never emitted, never consumed; scaffolded apps could only write filters that never match) and documented app:changed as a global app-list signal Node support (packages/sdk/src/node.ts): - the node transport ships a default websocket factory: global WebSocket on Node 22+, the ws package on supported Node 20 — bb.on and bb.data.onChange no longer throw out of the box on Node 20 - BbRealtimeSocket is now a minimal runtime-agnostic shape; default factories adapt the environment socket (browser global, node global, ws) instead of requiring DOM event types API surface cleanup: - *RealtimeOnInput types renamed to *RealtimeOnArgs and on(input) to on(args), matching the SDK-wide Args convention (pre-release rename) - BbSdk and InjectedAppWindowBb derive on() from the shared BbRealtime interface instead of triplicating the signature - CreateCurrentAppDataAreaArgs requires apps/realtime (the only caller always passed both); deleted the unreachable onChange runtime throw - shared cloneJsonValue helper replaces two diverged copies; app-data events are cloned once per delivery instead of twice - realtime URL derivation preserves a path-prefixed baseUrl (mirrors the HTTP transport); same-origin browser derivation still uses /ws - resolveApplicationId reuses requireCurrentApplicationId Frontend: - removed the unreachable app:changed cache registry (the SPA never subscribes to the app entity; system:apps-changed remains the canonical app-list invalidation path) and replaced the bypass test with one pinning the no-op - ws.ts logs dropped inbound messages instead of swallowing them Tests (+54): buffered-flush deliver/stale-path cases, negative id and prefix scope filtering, all previously unexercised dispatch paths, listener exception isolation, unsubscribe-during-dispatch/replay, backoff growth/cap/reset, close-during-pending-backoff, resync-on- reconnect ordering, late connection-observer snapshot, lenient parsing, hub fail-safe validation drop, notifySystem/notifyHost delivery, full change-kind schema-gate sweep per entity, and notifyGlobalAppsChanged end-to-end through real websockets. Known debt (deliberately not addressed here): BbRealtimeClient and the SPA's WebSocketManager remain two parallel realtime implementations; consolidating them is a standalone refactor. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
…ider turn watchdog
The idle watchdog killed healthy workflow turns ("No provider activity
for 906s after item/completed"): item/backgroundTask/progress was added
to the activity list, but those events are thread-scoped (turn_id NULL)
and the candidate query correlated activity strictly on the anchor
row's turn_id, so a streaming workflow was invisible to it.
The anchor is now the newest event that is either scoped to the active
turn (the latest turn/started) or a thread-scoped background task
event. The NULL-turn arm is deliberately restricted to the
backgroundTask family so thread-scoped provider/error noise cannot
defer reaping a wedged turn. item/backgroundTask/completed joins the
activity list — load-bearing, because progress rows are pruned the
moment the completed row lands, which would otherwise false-fire the
watchdog right after a successful workflow.
Guard hardening from adversarial review: active threads with no
turn/started yet are excluded in SQL (a NULL activeTurnId would throw
in row parsing and abort the whole sweep tick); the turn/completed,
pending-interaction, and started-at correlations all key off the
active-turn subquery instead of the anchor's turn_id; OR fragments are
self-parenthesized (drizzle's and() does not wrap raw fragments);
empty-string providerThreadId anchors fall back to the latest real id;
and the persisted lastActivityEventType is a plain string so
activity-list edits never make stored watchdog events unparseable.
The regenerated SDK browser bundle also catches up on the workflow and
reasoning-level domain schemas from a0bd295a7, which shipped without a
bundle regen.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
…s on disk
Rebuilding an app (or editing public/ files directly) never reloaded the
open app surface. Two breaks upstream of the SPA's existing reloadToken
mechanism: the daemon's apps-root watcher silently dropped every path
under <app>/public/ (no classification branch), and even reported storage
hints were gated behind an app-list signature diff that content edits
never alter.
Add a per-app content-changed signal end to end:
- host-watcher: classify <app>/public/** as application-content-changed,
deduped to one event per app per flush batch; source/ stays
unclassified (nothing served changes until a build writes public/).
- host-daemon: dispatch the new observed kind and send
{type: "application-content-changed", applicationId} over the daemon
WS, with per-app offline buffering flushed after reconnect.
- domain/server: APP_CHANGE_KINDS gains "content-changed"; app changed
messages carry an optional id (absent = list-level, present =
app-scoped); hub.notifyAppContentChanged broadcasts it, and notifyApp
is narrowed to the new AppListChangeKind subset so an id-less
content-changed is unrepresentable at the producer.
- SPA: subscribe to the "app" entity and invalidate only the changed
app's detail + markdown-preview queries via REALTIME_APP_CHANGE_REGISTRY,
so the detail refetch bumps dataUpdatedAt and busts the iframe
reloadToken without reloading other open apps.
- sdk: document the new app:changed semantics; regenerate the scaffold
bb-sdk.d.ts and browser runtime bundle.
Also extract the shared mock hub socket test helper that was previously
copy-pasted across four server test files.
Validated end to end against a live dev instance: touching
public/index.html broadcasts {entity:"app",id,changes:["content-changed"]}
within the watcher debounce; touching source/ broadcasts nothing.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
…ntime already hosts A thread.start/turn.submit whose freshly scanned injected-skill catalog hash differed from the loaded runtime's hash always forced a runtime replacement, which threw whenever the environment had active threads or open terminals. An agent that installed a skill mid-turn (e.g. building an app with a skills/ dir) therefore bricked its own thread: every subsequent message failed with "already has an active runtime with injected skill catalog X; requested Y" and was dropped, and the thread flipped to error status while the daemon-side turn kept running. Thread commands now pass their targetThreadId down to ensureCompatibleEntry; when the entry already hosts that thread and has active runtime work, the stale catalog is reused (with a warn log) and the refresh is deferred to the next launch on an idle environment. Idle entries are still replaced as before. The remaining conflict case (environment busy with other threads or terminals) now throws a typed SkillCatalogConflictError mapped to the new workspace resolution failure code "skill_catalog_conflict" instead of surfacing as "unknown". Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
…ersarial review Follow-up to the stale-catalog deferral fix, addressing review findings: - Keep the about-to-be-active catalog when pruning staging dirs during a runtime replacement. The cleanup keep-list was built only from loaded entries after the replaced entry was deleted, so the freshly staged new catalog was removed before the replacement runtime bound it — the refreshed runtime launched with skill roots pointing at a deleted directory (pre-existing, but the deferral fix made idle replacement the designed refresh path). - Defer the catalog swap for any busy runtime a thread command targets, not just runtimes already hosting the thread. Keying the guard on hosting left the original brick reachable: a terminal-first entry (no catalog, open terminal) or a sibling thread resuming after daemon restart still failed and dropped the message. SkillCatalogConflictError now only guards the no-target invariant. - Warn once per requested catalog per entry instead of on every command while the environment stays busy. - Stop CreateEntryArgs advertising ignored resolution-time fields (injectedSkillSources, targetThreadId). - Pin the full defer-then-refresh-once-idle sequence and staged-catalog survival in tests; the prior reuse test compared an entry's hash with itself. - Regenerate the SDK browser bundle for the skill_catalog_conflict contract code added in the previous commit. Known follow-ups (not in this change): the deferral is invisible to the server (no catalog hash in command results, no idle-transition reconciliation owner), and thread commands remain unlaned so a catalog-bearing command racing an in-flight resume can still replace the runtime mid-resume. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
…ture The unarchive-resume fixture landed on main without the new required turn.submit option introduced by the workflows commit; align it with the sibling fixtures. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
plans/claude-code-workflows.md was swept into the manager-templates removal commit by an over-broad add during conflict resolution. Plans live in thread storage, not the repo. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
…t, honest contracts, d.ts codegen Adversarial-review fixes for the apps platform surface: - extract scaffold machinery out of routes/apps.ts into services modules; fs.cp filtered copy applies exclusions at every depth; build-time template copy reuses it so dist ships no dev artifacts; README $-pattern safe - serve the window.bb runtime from a content-hashed immutable endpoint instead of inlining ~865KB into every app HTML response; bundle is regenerated from the real module graph (no regex+Function eval) with a drift-guard test against the committed artifact - delete decorative AppRuntimeBootstrap capabilities/dataUrl/messageUrl; appId/applicationId required on the injected window.bb contract - app d.ts codegen emits BbRealtime (window.bb.on typechecks in new apps), self-containment guards actually fire, drift test typechecks generated output; template source typechecked, source<->public drift guard, useTodos stale-snapshot race fixed, todo helper renamed for accuracy Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
…e entry - delete the ~240-line dead CLI HTTP client; port its tests to the SDK transport that production actually runs - typed BbHttpError (status + server code) replaces 'HTTP 404:' message sniffing; missing app vs missing data path report correctly again - importing @bb/sdk/node no longer eagerly loads CLI config — bb --help/ --version/guide work without BB_SERVER_URL; one canonical package entry (package.json '.' and vitest alias agree) - status area uses domain unions instead of plain strings; dead area exports removed; CLI consumers (guide, pending-todos, context-env) stop shadowing SDK types and building network clients for local work Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
- fix stale 'connected' observer state when the last targeted listener unsubscribes while the socket is CLOSING; fix reentrant connect during the 'disconnected' emit corrupting reconnectDelayMs; regression tests for both plus the synchronous-throw path through on()/activateListener - collapse four copy-pasted dispatch methods and eight listener-record interfaces into one generic keyed implementation; named aliases replace inline Extract<> signatures - hub.notifyApp signature matches reality (single change kind); hub tests assert distinct outcomes per kind; integration test reuses the SDK ws adapter; lenient/strict server-message schema drift guard; stale daemon-protocol comment removed Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
…watchdog + reasoning-ladder single source - lost background tasks now settle on every realistic path: daemon restart (regardless of previous-session status), lease expiry, and disconnect grace — mirroring pending-interaction reconciliation; settling no longer flips already-completed workflows to interrupted - late thread-scoped backgroundTask events no longer stretch the spawning turn's source range (turn-summary expansion 400 fix) with projection regression tests; error-only workflow rows are expandable; shared settled-state predicate and timeline row types deduped - backgroundTask progress rows fetched by targeted query instead of loading all rows and filtering in JS; timeline window backfill covered by tests and reuses mergeStoredEventRowsById - workflowsEnabled stays required downstream of the server boundary and the default policy is tested with true; watchdog thread-scope list derived from threadOnlyThreadEventTypes; SQL fragments deduped; custom-model reasoning ladders reuse the server policy table; dead taskType contract carry removed Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
- delete the dead first-boot PREFERENCES.md host read (and its mode plumbing through thread-send/queued-messages/nudge sweep) that could never find content after seeding removal; drop the dead request field - historical host.list_manager_templates command rows stay in the read-only prune cohort (test on in-memory sqlite) - inline the pass-through ManagerSlot wrapper; drop the stale 'may already exist from user templates' welcome-template sentence and the test pinning it Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
…at can fail - browser-view layout projected in a consistent coordinate space so docked DevTools no longer misprojects the native view; will-resize uses the in-flight bounds instead of stale getContentBounds() - attach/setBounds accept the legacy bounds shape under shell/server version skew; one shared collapsed-layout sentinel in the contract; open-tab payload uses the contract type; dead clamp export removed - BrowserTabDeck resize test flushes rAF and fails on revert; ThreadActionsMenu test exercises the actual no-drag fix; no-drag class gated on usesDesktopChrome Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
…ering - remove unreachable skill_catalog_conflict contract code; thread-brick regression test drives the real command path (targetThreadId from a thread command); staging-survival assertions check the real catalog hash instead of masking null with ''; log-throttle state moved off the exported RuntimeEntry - ServerConnection's three parallel recoverable-message buffers unified behind one keyed pending buffer (existing tests pin behavior) Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
Daily ship of
sawyer-next→ main: the 15-commit delta sincesawyer/ship-2026-06-03(#59), rebased onto current main, plus a full adversarial multi-agent review of the entire diff with all confirmed findings fixed.Features
@bb/sdkcore package,window.bbwith full CLI parity, CLI rewired as an SDK consumerbb.on(server hub → SDK subscriptions) + hardeningconfig.json(per-provider reasoning ladders).d.tsships in the packaged apppublic/changes; global-app skills catalog hardeningAdversarial review → fixes
A 238-agent review (11 units × 3 lens bundles, refuter panels per finding, completeness critic) confirmed 90 findings (10 high / 33 medium / 47 low; 20 refuted). All were fixed in the seven
fix(review):commits, including:workflowsEnabledrequired downstream of the boundary; deadtaskTypecarry removed; IPC layout change tolerates legacy shape under version skewwindow.bbruntime served from a content-hashed immutable endpoint instead of ~865KB inlined per app HTML load; backgroundTask rows fetched by targeted querybb-sdk.d.tsis valid (and the drift test now typechecks output); committed runtime bundle has a staleness guardDeliberate deferrals / follow-ups
buildTimelineTurnSummaryDetailsdoesn't backfill latest backgroundTask state for its range;selectAcceptedClientRequestContextRowscomputesafterSequencefrom post-injection rowsservices/apps/atomically (currently co-located inservices/threads/beside the template)provider.listValidation
turbo run build/typecheck— green across all 31 tasksturbo run test --force— 30/30 packages green (includes new regression tests for every behavioral fix)🤖 Generated with Claude Code