Conversation
… ram/threads (#10605) Co-authored-by: Shaw <shaw@elizalabs.ai> Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
… DB (#10730) Co-authored-by: Shaw <shawgotbags@gmail.com> Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
release: promote develop → production (money-integrity wave, monetization rails, auth overhaul, live rate limiting)
…2400s install timeout (#10839) (#11314) Prod deploys run from main (production environment allows only the main branch), but main cloud-cf-deploy still had the install-hang config: bun canary + install cache under $PWD (box /tmp, slow FS) + 900s per-attempt cap. Every money-integrity-wave prod deploy hung on bun install. Ports the three fixes already merged to develop (#11235/#11268/#11304): pin bun to latest, put the install cache on local $HOME (persistent/warm), and raise the per-attempt timeout to 2400s so the slow-but-completing install finishes. With this on main, a main-push deploy installs, then migrate-db runs (branch=main satisfies the production environment policy) pending the required-reviewer approval, then Worker/Pages deploy. Ships the merged money fixes (#11190 payout gate, #11189 cron, escrow, #11163 launch gate) to prod.
…#11428 conflict — same deploy fixes + migrate-db)
release: promote develop → main (money-integrity + fleet wave, 2026-07-02)
release: promote develop→main — ship #11413 money-gate + runtime restores to prod
release: promote develop→main — ship #11413 restores + migrate-runner fix to prod
chore(promote): develop → main — 36 commits (money fail-closed pricing, MCP debit refund, credential-pool + agent-api hardening, #11640 deploy-concurrency fix)
chore(release): promote develop to main — credit-gated app provisioning + money-surface cleanup
…e iOS load path (#11612) The plugin-local-inference LocalAIManager is stubbed out of the mobile agent bundle — the load that actually jetsammed the iPhone 16 Pro Max runs through LlamaBridgeImpl.swift (n_gpu_layers 999 + n_ubatch 1024 via mobileBatchSizes) and the @elizaos/capacitor-llama JS adapter. Apply the same admission fix there: - LlamaBridgeImpl.mobileBatchSizes: device physical micro-batch 1024 -> 256. Compute buffer 1037 -> ~260 MiB, so weights 4722 + compute 260 + KV 36 ≈ 5018 MiB fits the ~5461 MiB (2/3 of 8 GiB) jetsam working set with full-GPU speed retained. - LlamaBridgeImpl.loadModel: footprint model now includes the explicit n_ubatch-scaled compute-buffer term; admission order is ctx-shrink -> n_gpu_layers reduction (exact layer count via a CPU-only mmap metadata probe + llama_model_n_layer, exported by the shipped LlamaCpp.xcframework) -> clean failure. CPU-resident layers stay mmap-backed clean pages, so partial offload genuinely lowers the wired footprint; CPU-only loads no longer count un-wired weights against the budget. - capacitor-llama adapter: n_ubatch 512 -> 256 on the non-speculative mobile path; unload-on-failure — initContext failure releases the native context, and a failed decode (e.g. Metal ret=-3) unloads the model instead of leaving multi-GiB wired buffers mapped until jetsam. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
…-login-verification # Conflicts: # packages/cloud/api/__tests__/advertising-account-admin-routes.test.ts # packages/cloud/shared/src/db/migrations/meta/_journal.json # packages/cloud/shared/src/db/schemas/ad-accounts.ts # packages/ui/src/components/pages/__e2e__/output-launcher/desktop-rest.png # packages/ui/src/components/pages/__e2e__/output-launcher/launcher-walkthrough.webm # packages/ui/src/components/pages/__e2e__/output-launcher/mobile-after-swipe-home.png # packages/ui/src/components/pages/__e2e__/output-launcher/mobile-rest.png # packages/ui/src/components/shell/__e2e__/output-home/01-mobile-home.png # packages/ui/src/components/shell/__e2e__/output-home/02-mobile-notification-sheet.png # packages/ui/src/components/shell/__e2e__/output-home/03-mobile-launcher.png # packages/ui/src/components/shell/__e2e__/output-home/04-mobile-launcher-single-page.png # packages/ui/src/components/shell/__e2e__/output-home/05-desktop-home.png # packages/ui/src/components/shell/__e2e__/output-home/07-desktop-edge-buttons-home.png # packages/ui/src/components/shell/__e2e__/output-home/08-desktop-edge-buttons-launcher.png # packages/ui/src/components/shell/__e2e__/output-home/09-desktop-notification-panel.png # packages/ui/src/components/shell/__e2e__/output-home/home-screen.html # packages/ui/src/components/shell/__e2e__/output-home/mobile-launcher-flow.webm # packages/ui/src/components/shell/__e2e__/run-home-screen-e2e.mjs
…root cause + fixed-build load Old build reproduced #11612 as an empty reply thread (silent bf16 mul_mm failure, no GPU/crash on sim); STT + vision driven through the UI; the rebuilt bf16 MSL-3.1 + recoverable + OOM-admission slice installs/launches/loads without crashing. Completed sim reply not captured (sim translated-Metal pipeline compile is pathologically slow — env characteristic, not a product defect; real inference proven on Mac fused lib + device bf16-loads). Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
…aths Adds testCloudOnboardingChatAndVoice + testLocalOnboardingChatAndVoice to the device capture harness: completes the real in-chat first-run conductor, sends a prompt + awaits reply (chat), taps mic to assert recording state (voice, with SpringBoard permission-grant). Filmstrip + AX dumps; honest skips on AX variance / cloud-session-gated OAuth. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
…tal3.1 wedges the sim) The #11612 bf16 fix bumped all iOS metallib slices to metal3.1 so the A-series device slices emit kernel_mul_mm_bf16_f32. But the simulator's translated Metal toolchain wedges compiling that metal3.1 bf16 kernel set at model load (fixed slice pegs a core 20+ min with flat RSS, never finishes; the pre-fix 2.4 slice loaded fine). The simulator doesn't need embedded bf16 kernels — the runtime bf16-library gate (ggml_metal_device_init) routes bf16 ops to CPU when the kernel is absent — so generation still runs. Gate metal3.1 to the physical-device slices only; simulator slices revert to ios-metal2.4. Override via ELIZA_IOS_SIM_METAL_STD. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
…, wedge is in the #11612 fixed-build load path (likely OOM-admission probe) Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
…ted behind listLocalAgentBackups) The first-run greeting+placement choice is seeded only after client.listLocalAgentBackups() resolves, which waits on the agent API — on a fresh device the local engine is still waking, so 30s was too short. Poll up to the agent-ready budget (≤300s) for the placement choice before skipping. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
…) + document WKWebView XCUITest-invisibility blocker (#11612 device reply) The on-device local-onboarding capture is blocked because the Capacitor WKWebView onboarding controls are not exposed to XCUITest's element tree — the placement choice renders on screen (evidence screenshot) but no button/staticText is queryable. Hardened tapWebChoice with substring matching + a coordinate-tap fallback on a matching staticText (helps any exposed web content) but the deeper fix is app-level WKWebView accessibility. Real-device tap injection has no other route (idb is sim-only; pymobiledevice3 has no touch). Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
…d + locked-composer placeholder The onboarding-to-home ui-smoke suite (10 specs, desktop + mobile) had been fully red: the shared helper still asserted the removed runtime:other choice (the chooser is now cloud/local/remote — #11509) and the old locked-composer placeholder 'Choose an option to continue' (shipped text is 'Tap a highlighted option above to continue'). Point the runtime assertion at 'remote', reach the provider:other handoff via runtime local, and match the real placeholder. All 10 pass — cloud + local + cloud-inference + other-provider + remote + tutorial, desktop and mobile touch. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
…GPU-OOM fix loads gemma4-2b on A18 (0/36 GPU offload, no OOM) BootCaptureUITests: raw normalized-coordinate taps for the invisible WKWebView onboarding rows (placement + provider) + a re-send-until-ready chat loop. On-device ggml.log proves the #11612 memory-admission fix: eliza-1-2b (gemma4, 601 tensors) loads with n_gpu_layers reduced to 0/36 to fit the A18 working set — no kIOGPUCommandBufferCallbackErrorOutOfMemory, no ret=-3, no jetsam (pre-fix OOM'd at 999). Final generated-reply text pending one run with iOS Enable-UI-Automation active (re-prompts passcode after idle — iOS security gate). Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
… to stop the auto-lock stall (#11841) On iOS the on-device runtime streams the multi-GB Eliza-1 model download in-process (JSContext compat / sideload path). When the device auto-locks mid-transfer the runtime is suspended and the download stalls at "Loading eliza-1-2B…". This wires a reference-counted `keep_awake_set` host function on the `__ELIZA_BRIDGE__` compatibility bridge so the downloader disables `UIApplication.shared.isIdleTimerDisabled` for the duration of an active transfer and restores it in the job's finally (balanced acquire/release; overlapping downloads compose). - KeepAwakeBridge.swift: new reference-counted idle-timer bridge module, installed from BridgeInstaller into the JSContext and held on BridgeKit. - downloader.ts: setDownloadKeepAwake(true) on entering "downloading", (false) in the runJob finally — swallowed best-effort, never fails a download; safe no-op off-device and on the full-Bun engine. - downloader.test.ts: 3 real tests over the download harness — releases on completion, releases on failure (balanced), and a throwing bridge never breaks the transfer. Full suite 20/20 green; typecheck + biome clean. Scope: this removes the common *auto-lock* stall on the sideload/JSContext path. A *manual* lock or backgrounding still needs the native background URLSession download — the tracked #11841 primary fix. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
the manual redirect loop rebuilt each hop init from params.init, re-sending method+body on every hop. 624325a (#11693) stripped credential headers cross-origin but left the body: a POST with a secret body that a compromised server 302s to an attacker origin re-POSTed the full body there. it also deviated from standard fetch, which the header fix cites: 301/302-POST and any 303 rewrite to a bodyless GET. reproduce that rewrite (whatwg http-redirect fetch): 301/302 on POST and 303 on any non-GET/HEAD method switch to GET, drop the body, and delete the body-describing headers (content-type/length/...); 307/308 still preserve method+body. fixes both the cross-origin body exfil and the GET-after-303 functional deviation. (cherry picked from commit 70e5c35)
…ub-agents via OPENCODE_CONFIG_CONTENT gateway mode (86bc107, #11651) promised a child env dump contains no raw provider key, but an opencode spawn broke the invariant twice over: buildOpencodeSpawnConfig embedded the raw cerebras/opencode/cloud key as provider options.apiKey inside the JSON assigned to env.OPENCODE_CONFIG_CONTENT, and applyModelGatewayEnv — which runs after and deletes only the seven named env keys — never touched it. the child env carried the raw key AND the opencode child pointed its baseURL straight at api.cerebras.ai / eliza cloud, bypassing the gateway. - buildOpencodeSpawnConfig checks resolveModelGatewayConfig() first, before any credential read (env, runtime settings, or config-env — setting() falls back to all three, so env-key deletion alone could never fix this). in gateway mode it returns a gateway-pointed openai-compatible provider config: nothing raw to leak, and the child routes through the gateway instead of bypassing it. - applyModelGatewayEnv enforces the stated invariant literally: raw values captured from the named keys are swept out of every remaining env value, so a composite carrier (a JSON blob like OPENCODE_CONFIG_CONTENT) that still embeds a raw key is dropped whole. fail-closed backstop for future merge steps. off-mode behavior is byte-identical; a new off-mode test pins the legacy direct-cerebras wiring. model-gateway-env.test.ts now exercises agentType opencode — the untested gap that hid this. (cherry picked from commit 3f6f49a158371957d2d50ffa8e723d5adbb471e6)
… agent-wake probe
Fresh-install onboarding stranded the user at a locked composer ('Tap a
highlighted option above to continue') with NO visible choices whenever the
local agent was slow/wedged to wake: the conductor only seeded the greeting +
runtime choice inside the then()/catch() of client.listLocalAgentBackups(),
which hits the still-booting agent API and can hang indefinitely.
Now the greeting + runtime choice seed synchronously on mount, independent of
agent readiness; the local-backup probe is a purely additive upgrade that
appends the restore choice below the greeting only if backups exist AND the
user hasn't already advanced past it (atomic updater, no double-seed). Two new
conductor tests cover the never-settling probe (greeting still appears + a
runtime pick works) and the racing-late-backups guard.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
… + add onboarding probe The assistant message bubble in ContinuousChatOverlay's ThreadLine wrapped its content in role=button aria-label="Show message actions" whenever it had row-actions. WebKit maps an ARIA button to an ATOMIC accessibility leaf (its aria-label becomes the node name and ALL descendants are dropped), so the onboarding greeting text and every ChoiceWidget button vanished from the native AX tree: unreachable by VoiceOver AND by on-device XCUITest (only the collapsed "Show message actions" node was exposed). Fix: a choice/form/followups-bearing assistant turn (every first-run onboarding turn) now stays a plain container -- bubbleInteractive gains a !hasInteractiveWidget guard (detected via the existing choice/form/followups parsers). Normal turns keep the tap-to-reveal Copy/Play/Edit row unchanged. Also adds an sr-only onboarding-state-probe (mirroring chat-detent-probe / home-launcher-page-probe) exposing the current first-run step id + choice ids so on-device automation can observe/drive onboarding deterministically. Two new unit tests cover both. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
…path (#11841) The earlier keep-awake fix (6ea36ad) only installed keep_awake_set on the JSContext compat bridge — but the production/sideload iOS build runs the full-Bun engine (confirmed on device: `compiledEngine=full-bun`, `[FullBunEngineHost] loaded ElizaBunEngine abi=3`), where that host function was unregistered and the downloader's call was a silent no-op. This routes it through the full-Bun host_call path so the idle-timer hold actually fires on the device that has the stall. - FullBunEngineHost.swift: add a `keep_awake_set` case to the host_call switch → KeepAwakeBridge.shared.setEnabled(enabled). - KeepAwakeBridge.swift: add a process-wide `shared` singleton + public `setEnabled(_:)` so both engine paths drive one ref-counted idle-timer holder; BridgeInstaller now installs the shared instance. - ios/bridge.ts: define `globalThis.__ELIZA_BRIDGE__.keep_awake_set` at boot (full-Bun) as a fire-and-forget `callIosHost("keep_awake_set", …)`, so the in-process downloader reaches native. Only defines it when the engine hasn't already; failures can't affect the download. typecheck (bridge.ts) + biome clean; KeepAwakeBridge parses clean; the native case uses existing helpers (boolValue/encodeHostEnvelope). Device verification (build → deploy → download completes with the timer held) follows. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
…ixes local onboarding "not found" loop
On a dev iOS device the local agent runs through the JSContext compatibility
kernel (ios-local-agent-kernel.ts), which implemented GET /api/first-run/status
but had NO POST /api/first-run handler. finishLocal's submitFirstRun POST fell
through to the catch-all 404 ("Not found"); the conductor's seedError turned
that into a re-offer of the "where should your agent run?" runtime chooser — the
on-device "local path → not found → pick the targets again" loop the user hit.
(The full-Bun store build + the mocked web e2e both serve the route, which is
why neither caught it.) Add the handler: accept + ack the finish payload
(no-op success matching the full-Bun bundle). Regression test asserts the
kernel serves it 200 instead of 404.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
…ngine (split JSC-free core) The full-Bun podspec source list is explicit and omits JavaScriptCore, so `bridge/KeepAwakeBridge.swift` (which imported JavaScriptCore) was never compiled there — `FullBunEngineHost` referencing `KeepAwakeBridge.shared` failed to build (`cannot find 'KeepAwakeBridge' in scope`, xcodebuild 65). Split it: the JavaScriptCore-free ref-counted idle-timer core stays in `KeepAwakeBridge.swift` (now added to the full-Bun source list), and the JSContext `install(into:)` closure moves to a compat-only `KeepAwakeBridge+JSContext.swift` (picked up by the compat glob, excluded from the full-Bun explicit list). Verified via `pod install` with ELIZA_IOS_FULL_BUN_ENGINE=1: KeepAwakeBridge.swift is now in the Pods project, the JSContext extension is not. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
…ume into chat after WebView eviction On a native device the first-run cloud login opens an external browser (SFSafariViewController) that backgrounds the WebView; iOS frequently cold-launches the app on return, wiping the conductor's in-memory flow state and the volatile __ELIZA_CLOUD_AUTH_TOKEN__ global. The result was a "restart": the user was bounced back to "where should your agent run?" instead of landing in chat. Two durable channels close the gap: - Persist the cloud token to the steward-session store (getCloudAuthToken reads it first) on poll/direct auth success, so elizaCloudConnected recomputes true after relaunch. - A cloud-login resume marker (localStorage) armed at cloud/hybrid pick time and cleared on completion or a fresh runtime re-pick. On relaunch the conductor rehydrates the interrupted flow (draft + pending-resume) and continues into chat instead of re-seeding the greeting. Fixes the effect-ordering trap where the auto-resume effect fired before the mount effect armed the marker: the resume body is extracted into runCloudResume and kicked directly when the connection is already live at mount. Tests: first-run-cloud-resume unit (round-trip/validation), conductor regressions (marker armed on pick, resume-on-relaunch with no greeting restart, cleared on fresh local pick). 26 pass; typecheck + biome clean. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Vq3VfLC939jy9e4aA12KGP
…sh first-run (MoonCycles A18) Fresh-install XCUITest capture: locked onboarding chooser → composer unlocked after the local finish path (POST /api/first-run) succeeds — no "not found → pick again" loop. Plus README documenting both onboarding fixes, the 26 passing tests, and the remaining manual cloud-OAuth step. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
…lanner and trajectories can tell siblings apart The agent picks actions by name + description (+ similes at Stage 1). Several siblings share nouns/similes with no "when NOT to use" guidance, and duplicate action/provider names were dropped silently at debug level. This lands the high-confidence, mechanism-level disambiguation: - routingHint is now the canonical "use when X; do NOT use when Y -> OTHER" carrier (documented in the Action type). Unlike `description` it is prepended verbatim to the planner tool (never compressed / 160-char truncated) and is captured in trajectories via the planner stage's model.tools. - Populate routingHint on the top cross-domain collisions: orchestrator TASKS (reciprocal to SCHEDULED_TASKS: reminders/routines -> SCHEDULED_TASKS/OWNER_*), WEB_SEARCH (external web only; private data -> MEMORY; a URL -> WEB_FETCH), MEMORY, USE_SKILL, MCP, MESSAGE (reply -> REPLY, channel membership -> ROOM, feed -> POST), OWNER_GOALS (goals vs reminders/todos/routines). - Prune generic cross-cutting similes: USE_CAPABILITY/RUN_CAPABILITY off USE_SKILL; USE_TOOL/RUN_TOOL/EXECUTE_TOOL/INVOKE_TOOL/GET_RESOURCE/… off MCP; ROOM off MESSAGE (collided with the ROOM action). - Resolve the live SHELL x2 collision: rename the agent terminal action SHELL -> TERMINAL_SHELL (the coding-tools SHELL is the fuller impl and already owns the generated SHELL spec), drop its EXEC/RUN_COMMAND/CALL_MCP_TOOL similes, add a routingHint, keep it in the terminal-context compaction keep-set. - Fix REPLY's description that pointed at connector send-actions that no longer exist -> MESSAGE(action=send)/POST (core.json + regenerated action-docs). - registerAction/registerProvider now warn (was debug) on a duplicate name so load-order-dependent collisions (SHELL, SETTINGS/ROLES/CHARACTER/MCP providers) surface at boot instead of silently first-wins-dropping. - Document OWNER_GOALS x2 as an intentional two-topology design (PA primary, plugin-goals fallback; same GoalsService) so it isn't "fixed" into a regression. Verified: typecheck clean (core, agent, mcp, agent-skills, orchestrator); prompts tests 34/34; action-docs regen scoped to the REPLY lines. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
…ive for the local model download Two fixes uncovered while proving the #11841 on-device model download: 1) Voice/command onboarding parser (first-run.ts): the on-device option is labelled "On this device (recommended)", but the inference-step parser's cloud regex matched "recommended" — so an explicit local pick ("on this device (recommended)") set localInference="cloud-inference" and the local model download never triggered. "recommended" is not a cloud signal; cloud options still match on cloud/eliza cloud/online, and a bare "recommended" falls through to the same cloud-inference default. Added a regression test asserting the labelled option yields all-local + firstRunDownloadsLocalModel === true. 2) On-device onboarding capture (BootCaptureUITests + ios-device-capture): the local model download is fired fire-and-forget by first-run finish, but the XCUITest tore the app down before the agent-wait + download completed, so the ~5GB pull never landed in prior capture runs. Added an opt-in holdForLocalModelDownload step (ELIZA_LOCAL_MODEL_DOWNLOAD_WAIT_SECONDS, default 0) that keeps the app foregrounded — touching the AX tree so XCUITest doesn't idle-kill it and filming the wait — so the download runs to completion; [Downloader] completion + the native keep_awake_set hold are verified from device syslog. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
…on on the tool stage The tool-execution trajectory stage (RecordedToolStage) recorded only name/args/result — a reviewer or training pipeline judging "did the planner have enough to disambiguate this action?" had to cross-reference the preceding planner stage's model.tools. Capture the exposed ToolDefinition.description (the action's routingHint + compressed description — exactly what the planner saw) on the execution record so it is self-contained. Sourced from the type-safe PlannerLoopParams.tools, not runtime.actions. Rendered in the markdown review as `- description:`. Test covers JSON round-trip + markdown render. Verified: core typecheck clean; trajectory-recorder 37/37; planner-loop 92/92. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
…on descriptions The action/provider description text shown to the planner was hard-truncated to 160 chars (compressPromptDescription) and the same cap was enforced as a lint rule (description-compressed-lint) — so any "use when / do NOT use when" clause past 160 chars was silently cut before the model saw it, and 22 baked descriptions in the generated docs were already clipped with "...". There is no reason to cap the text out to the model. This removes both caps: - compressPromptDescription now returns the full normalized/compressed text (no MAX_DESCRIPTION_LENGTH, no truncateDescription); code spans/URLs/paths still protected, filler still stripped, but nothing is cut. - lintDescriptionCompressed drops the length rule (keeps the filler/abbrev/ imperative-lead style rules). - Regenerated action-docs.ts + plugins.generated.json: all previously-truncated descriptionCompressed values are now full (remaining "..." are legit content: example texts, [[x,y],...] notation). Verified: prompts 34/34, core description-compressed-lint 10/10; core + prompts typecheck clean. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
evm/actions/transfer.ts exported a `transferAction` (name "WALLET") that was never registered — the live WALLET action is walletRouterAction, and this subaction's implementation is the TransferAction class + buildTransferDetails (both kept, still used by chains/registry.ts + evm/chain-handler.ts). Removing the orphan const (zero references repo-wide) and its now-unused imports; the live transfer path is untouched. Verified: tsc + biome clean; no remaining references. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
…unctuation ebd5203 (#11573) fixed the dead e.g/i.e entries by matching [\w.]+ but anchored the match with (?:^|\s), which rejects abbreviations preceded by quotes/parens/asterisks ('"Dr' / '(Mr') that the original \b handled — the tts first-sentence early-emit path chopped mid-name ('He cited "Dr.'). drop the anchor: leftmost matching already captures the maximal trailing [\w.] run and any other char or start-of-string delimits it. (cherry picked from commit 99511f2)
Contributor
|
Too many files changed for review. ( Bypass the limit by tagging |
Contributor
|
Important Review skippedAuto reviews are disabled on this repository. Please check the settings in the CodeRabbit UI or the ⚙️ Run configurationConfiguration used: Organization UI Review profile: CHILL Plan: Pro Run ID: You can disable this status message by setting the Use the checkbox below for a quick retry:
✨ Finishing Touches🧪 Generate unit tests (beta)
Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out. Comment |
Member
Author
|
Recommend: |
Member
Author
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
2 participants
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.
Accumulated worktree residuals from the shared swarm branch (its prior PR #11731 already merged; this batches the 42 commits since). Grouped by thread below. CI is the gate; individual threads are self-contained fixes.
Headline: iOS on-device model download (#11841)
keep_awake_setthat holds the iOS idle timer open while a model download is active, so auto-lock can't suspend the in-process transfer. Wired on both engine paths: the JSContext compat bridge and the full-Bunhost_calldispatch (FullBunEngineHost), with the JavaScriptCore-free core split so it compiles into the full-Bun target. Verified on-device: BUILD SUCCEEDED + installed on an iPhone 16 Pro Max running the full-Bun engine; thehost_callpath is live (host call method=…). Downloader unit tests 20/20.localInference=cloud-inferencebecause the inference-step parser's cloud regex matched the word "recommended", so the local model download was never requested. Removed "recommended" from the cloud signal (cloud still matches on cloud/eliza cloud/online); added a regression test (first-run 22/22).holdForLocalModelDownload(opt-inELIZA_LOCAL_MODEL_DOWNLOAD_WAIT_SECONDS) so the local-onboarding capture keeps the app alive through the fire-and-forget download instead of tearing it down; harness knob added.Other threads (self-contained)
Evidence
host_callliveness + 20/20 downloader tests attached to iOS on-device model download stalls when the app backgrounds / device locks (no background URLSession for the ~5GB pull) #11841. Live keep_awake_set-during-download capture is staged behind a CoreDevice tooling wedge on the build host (needs a host reboot) — N/A this batch, reproduce command in iOS on-device model download stalls when the app backgrounds / device locks (no background URLSession for the ~5GB pull) #11841.🤖 Generated with Claude Code