The Installed-Build Reliability Upgrade

This release focuses on keeping installed apps rock-solid—especially when loading key modules like search and memory. Onward and upward! 🎉

Highlights

Modules & installed builds 🧩

A targeted hotfix improves module admission in installed builds so critical modules load correctly again—unblocking both search (tinysearch) and memory (tinymemory) where they could previously fail due to module directory permission checks.

(#7155) — Thank you @senamakel!

Contributor Credits

Thanks to @senamakel for the fix and the fast turnaround!

Full Compare

v0.64.14...v0.64.15

The Files, Memory & Voice Power-Up

This release is a big step forward in day-to-day usability: clearer files and sessions, a major memory evolution, and real-time voice agents—plus a ton of reliability and performance work across the stack. 🎉

Highlights

Files & project workspace clarity 🗂️

Agent outputs now land in a visible Files folder, you can choose where that folder lives, and the app can surface those generated files directly in-place—making it much easier to find and use what your agent created. Large tool outputs are also handled more safely to keep projects tidy. (#6820, #6825, #6833, #6868, #6947) — Thank you @M3gA-Mind and @senamakel!

Memory v2, imports, and safer organization 🧠

Memory gets a major upgrade: engine-selectable memory v2, better lifecycle handling around each turn, and stronger portability when switching engines. The Memory UI and provider cards were modernized (including CortexDB), credits/free-period behavior is more consistent (with top-up prompts), and recall becomes time-zone aware.

Imports and migrations are significantly hardened: clearer “building vs empty” states, smarter batching/retries, better failure visibility with retry controls, more reliable resuming, and improved document ingestion (PDF/Office). Organization also improves with connector-first filing, better GitHub scoping, optional attribution to senders, and multiple erasure/deletion correctness and privacy/security fixes. (#6841, #6843, #6855, #6869, #6873, #6874, #6875, #6892, #6901, #6924, #6949, #6993, #7011, #7012, #7013, #7014, #7034, #7029, #7035, #7036, #7038, #7041, #7043, #7063, #7085, #7089, #7090, #7091, #7094, #7096, #7100, #7102, #7103, #7104, #7105, #7106, #7107, #7108, #7109, #7110, #7111, #7112, #7113, #7114, #7122, #7123, #7126, #7128, #7129, #7130, #7131, #7134, #7137, #7138, #7139, #7142, #7143, #7144, #7146, #7147, #7149) — Thank you @M3gA-Mind, @senamakel, @CodeGhost21, @YellowSnnowmann, and @lorenzozanee!

Live voice agents & voice foundations 🎙️

Voice agents are now truly live: talk to Tiny in real time with tool support (including Gemini Live, ElevenLabs, and Sarvam), plus a redesigned Voice agents experience with service cards, settings modals, and app-wide toasts. Under the hood, voice-related inference components were also reorganized for a cleaner foundation. (#7046, #7092, #6840) — Thank you @senamakel!

Chat UX upgrades & smoother conversation control 💬

Chat gets a substantial UI uplift: working folder support, per-model thinking controls, real thread titles, richer sidebar context, and clearer run-state cues. Reliability improves too—Stop is more resilient, streaming works even without a progress sink, reply-language directives are consistently sent, and several web chat delivery edge cases are fixed. The app also adds translatable error copy and improves transcript storage formats. (#7048, #7045, #6916, #7088, #7097, #7098, #6844, #7087, #6881, #6883) — Thank you @senamakel and @M3gA-Mind!

Agent intelligence, tool governance, and turn reliability 🧰

Agents are more robust and predictable: improved tool calling defaults (JSON auto) with a setting, better “finish” verification, fewer false failure classifications, clearer file-read behavior, and multiple fixes around context compaction, reasoning bounds, iteration caps, and hosted-path correctness/telemetry. Tool policy/approval gates were reworked, per-turn tool withholding was added, and timed-out commands now terminate their whole process group.

This also includes targeted fixes for web research budgets, credential scrubber behavior (including avoiding code redaction), handling HTTP status correctly, and improved orchestration behaviors. (#6910, #6909, #6963, #6965, #6966, #6967, #6968, #6969, #6971, #6972, #6973, #6974, #6975, #6976, #6977, #6980, #6992, #6889, #6856, #6879, #6884, #6945, #6946, #7117, #7118, #7119, #7120, #7140, #6865) — Thank you @senamakel and @sanil-23!

Sessions, scheduling, connectors & integrations 🧩

A new host-injected session store moves...

The Speed, Structure & Reliability Upgrade

This release tightens the core experience from end to end—faster first turns, clearer chat feedback, sturdier integrations, and a big internal modularity push that sets us up for even quicker iteration. 🎉

Highlights

Performance & responsiveness ⚡

The agent loop is noticeably snappier: first-turn prompts are dramatically smaller, prompt caching stays warm (improving time-to-first-token), and the managed inference path avoids unnecessary per-call client setup. Tool-output compaction is also enabled by default again—with new “REPL-style” tools over stored results to keep large tool outputs from stalling turns.

(#6787, #6788, #6823) — Thank you @senamakel!

Chat UI & approvals polish 💬

Chat got a round of UX improvements: long source lists now collapse cleanly (with better favicons), web-search rendering no longer leaves awkward gaps, and the silence watchdog now warns instead of killing a running turn—staying in sync with the core’s actual turn state. Approvals are also smoother: answered approvals won’t “ghost” into the background deck, and refusals now carry a clear deny reason that the model can actually read.

(#6802, #6827, #6800, #6810) — Thank you @senamakel and @M3gA-Mind!

Memory & knowledge engines 🧠

Memory is now explicitly selectable: you can choose exactly one active engine, including Local (TinyCortex) or CortexDB via TinyHumans, plus other supported providers. Switching engines can also copy existing memories across—making experimentation and migration much less painful.

(#6777) — Thank you @senamakel!

MCP & Composio integrations 🔌

MCP tooling becomes much more ergonomic and robust: server tools are exposed as stable, individually-addressable tools from a persistent cache, and string-encoded JSON arguments are normalized so calls don’t fail unexpectedly. Refusals from tool listing now tell the model what it can do (and avoid pointing at RPC-only methods). On the Composio side, embedders can give each agent its own credentials, and backend-routed Composio calls now flow through the tinyconnectors module instead of an in-process client.

(#6767, #6769, #6818, #6689, #6796) — Thank you @senamakel, @M3gA-Mind, and @oxoxDev!

Security, safety & policy hardening 🛡️

Security protections are sharper and less noisy: the prompt-injection guard is less likely to false-positive on recalled transcripts, and browser IP blocking now uses the full non-global range definitions (with regression tests). Shell execution reliability also improves in scheduled jobs on Windows, and shell scanning logic is consolidated upstream for consistency.

(#6776, #6709, #6764, #6790) — Thank you @senamakel, @Avi-Bendetsky, and @sanil-23!

Architecture & vendor modularity push 🧱

A major theme of this series is extracting host-independent logic out of openhuman-core and into the appropriate vendor/tiny* modules—keeping core focused on product policy, adapters, and contracts. This includes splitting RPC/server responsibilities, moving channels/skills/web3/filesystem/hooks/session/runtime/tooling helpers into their owning crates, and updating vendor pins to merged upstream mains. Along the way, stale subsystem references were cleaned up, and the tool-call type is now adopted directly from tinytools to reduce duplication and drift.

(#6703, #6779, #6780, #6782, #6783, #6784, #6785, #6786, #6789, #6791, #6793, #6794, #6795, #6797, #6804, #6805, #6806, #6807, #6809, #6773) — Thank you @senamakel!

Reliability, updates & CI/e2e stability ✅

This set also focuses on “keeping main green”: E2E web runs now avoid port collisions and refuse taken ports, the core updater once again resolves the correct versioned archive and stages the real binary, and managed backend outages now trigger bounded exponential backoff with explicit terminal outcomes. CI coverage improves too, including executing module-gated Rust tests, better layout gating behavior, and keeping vendor/lockfile pins aligned (including tinymcp and tinymemory pin handling).

(#6719, #6770, #6569, #6605, #6817, #6771, #6772, #6798) — Thank you @obchain, @lorenzozanee, @M3gA-Mind, @sanil-23, and @senamakel!

Desktop, voice & platform fixes 🖥️

TinyComputer gains saved flows, task reports, and a live Bali booking E2E to make real-world debugging and iteration much easier. There are also key platform fixes: macOS desktop builds were restored after upstream accessibility API changes, and vendor bumps include a Windows build fix for unsafe_code constraints in tinyagents.

(#6761, #6822, #6839) — Thank you @senamakel and @M3gA-Mind!

Web3 & x402 improvements 💱

Web3 functionality continues its move into tinywallet, with follow-on pinning to pick up important x402, fee, probing, and chain-status fixes—plus host-side support and tests to keep behavior correct and traceable.

(#6792) — Thank you @senamakel!

Usage & accounting accuracy 📊

The Usage page now reflects the spend it actually recorded: workspace tracking is fixed across auth transitions, duplicate recordings are avoided (and legacy duplicates ignored), and tiny UI details like <$0.01 and negative-zero rendering are corrected. The “Token & Cost” tab was also renamed to “Token savings” in all locales to match what it truly shows.

(#6720) — Thank you @graycyrus!

Embedder & agent control knobs 🎛️

Embedders now have tighter control over runs with a per-turn tool-call budget limiter, without needing to alter persistent agent configuration. There’s also cleanup after the inline-skills/deferred-tools work: vestigial spawn args are removed and e2e tests fixed accordingly.

(#6350, #6762) — Thank you @theamazinghenk and @senamakel!

Settings & provider catalog changes 🔁

A built-in Cheaper Inference provider entry was added, then reverted—returning settings and documentation to their prior routing behavior.

(#6620, #6778) — Thank you @aiapienthusiast and @senamakel!

Docs & developer workflow helpers 🧰

Internal contributor tooling improves with a new tracked Claude skill guide for migrating code into vendor submodules, and tinyskills is bumped to fix sourceUrl resolution behavior for non-SKILL.md Markdown blobs.

(#6775, #6819) — Thank you @senamakel and @M3gA-Mind!

Release housekeeping 🚀

This range includes the release PR used to cut and ship the bundle.

(#6826) — Thank you @M3gA-Mind!

Additional highlights 🔗

A few more focused contributions round out this release with targeted fixes and improvements that are worth calling out. (#6781, #6803) — Thank you @senamakel!

Contributor Credits

Thanks to everyone who contributed to this release: @aiapienthusiast, @Avi-Bendetsky, @graycyrus, @lorenzozanee, @M3gA-Mind, @obchain, @oxoxDev...

The Intelligence Upgrade 🎉

v0.64.4 to v0.64.7 brings 12 PRs across 12 commits, with upgrades across memory, voice, agents, reliability, and developer foundations. Thank you to everyone who contributed to this release. ✨

Highlights 🚀

✅ Tasks, chat & agent workflows

Task, chat, and agent workflows become easier to steer and inspect, with better task boards, persistent run state, safer delegation, and clearer model/thread controls. (#6672, #6683, #6702, #6760) — Thank you @M3gA-Mind and @senamakel!

🧩 Integrations, UI & developer foundations

Integrations, UI polish, and developer foundations also move forward, including clearer connected-account surfaces, meeting-agent wiring, presentation/artifact improvements, and cleaner docs/code organization. (#6668, #6678, #6693, #6699, #6700, #6701, #6704, #6759) — Thank you @senamakel, @M3gA-Mind, and @siddarthkay!

Contributor Credits 🙌

• M3gA-Mind (#6672, #6693) — thank you!
• senamakel (#6668, #6678, #6683, #6700, #6701, #6702, #6704, #6759, #6760) — thank you!
• siddarthkay (#6699) — thank you!

Full Compare 🧭

v0.64.4...v0.64.7

The Better Windows Desktop Upgrade

This release is all about making the Windows desktop experience more reliable, more polished, and smoother to operate—especially when you’re offline. 🎉

Highlights

Offline-ready Windows installers 📦

Windows desktop installers now bundle all registry-pinned native modules (and their extracted DLLs) directly into the installer, so key features don’t depend on a first-run download. Module resolution now prefers the read-only installer bundle before user cache or network, and the Windows checks/smoke steps were updated to verify offline module loading end-to-end. (#6687) — Thank you @senamakel!

Sharper Windows window controls & cleaner shutdown 🪟

Windows gets a more compact first-launch window and improved custom window controls (minimize, maximize/restore, close), with the drag strip staying available during boot and error screens. Shutdown behavior is also fixed so closing properly exits both the desktop host and embedded core (including support for the app_quit IPC command), plus the dev launcher was simplified with earlier architecture mismatch detection. (#6692) — Thank you @stevent95-star!

Contributor Credits

Thanks to everyone who contributed to this release: @senamakel, @stevent95-star.

Full Compare

v0.64.3...v0.64.4

The Reliability & Speed Upgrade

v0.64.0 → v0.64.3 is all about making OpenHuman feel steadier under real workloads—fewer loops, better recovery, faster turn startup, and clearer billing/usage surfaces. Ship it. 🚀

Highlights

Agent stability & loop control 🧭

Agents now stay focused on the latest user request (even after tool search results arrive), avoid getting pulled back into earlier greetings, and stop unproductive tool loops sooner with smarter failure classification and bounded exploration. The result: fewer stalls, fewer repeat calls, and more deterministic “partial-but-useful” outcomes when tools are blocked. (#6656, #6660, #6661) — Thank you @senamakel!

Chat recovery, streaming resilience & warm-session correctness 🛠️

This release hardens the chat experience: warm-session completions now properly finalize progress and can optionally capture full successful inference SSE responses for debugging; completed model streams that return no answer get a bounded one-time retry; and transcript/app-state recovery is more robust after failed or torn writes—preserving history and keeping queued follow-ups stable. (#6665, #6685, #6686) — Thank you @senamakel!

Faster integrations & skill discovery 🔥

Turn startup and tool availability should feel snappier thanks to caching and warming of Composio integrations and skill metadata across turns/sessions (with clear invalidation rules). This also forwards Socket.IO WebSocket upgrades through the local inference capture proxy for more consistent event routing. (#6684) — Thank you @senamakel!

Time, time zones & scheduling correctness 🕒

OpenHuman now understands more natural time phrases (like “tomorrow 9am”, “since Monday”, and “in 2h30m”), and Composio connector calls include the user’s IANA time zone—helping downstream results (like Gmail timestamps) line up with what users actually mean. Durable subagents also resume more reliably when session IDs show up in multiple fields. (#6670, #6673, #6674) — Thank you @senamakel and @M3gA-Mind!

Observability & developer telemetry 📡

Production Langfuse tracing is enabled from the canonical production API host (while keeping unknown hosts closed), and child agent turns are grouped cleanly under parent conversation lineage—with more complete generation metadata and careful accounting to avoid double counting. (#6677) — Thank you @senamakel!

Clearer billing & usage UI 💳

Usage and billing surfaces got a reality check: “Usage & Limits” is now “Usage” (removing misleading local budget gauges), provider-call usage logging has a dedicated tab with better filters and bounded record semantics, and billing access is consolidated under Accounts with a hosted dashboard link (with /settings/billing preserved as a redirect). Also, chat no longer incorrectly blames empty model responses on insufficient credits—credit guidance now appears only for explicit billing errors. (#6657, #6679) — Thank you @senamakel!

Tooling & platform cleanup (Windows + retired components) 🧹

Windows users should see more reliable TinyMemory module loading thanks to a pinned TinyBus ACL fix, and the retired Medulla integration has been fully removed (while preserving a legacy config alias so older configs load safely as local). The orchestrator tool surface was also tightened by removing unused prompt tools and delegate entries. Finally, file_read is now unpacked so the orchestrator can read result previews without opening up the whole files tool pack. (#6680, #6682, #6669, #6666) — Thank you @senamakel and @M3gA-Mind!

Test coverage & regressions ✅

New regression coverage locks in the “greeting → deferred Gmail tool continuation” behavior end-to-end, helping ensure the agent keeps the active request and tool result together and doesn’t fall back into earlier context. (#6659) — Thank you @senamakel!

Contributor Credits

Big thanks to everyone who shipped v0.64.0 → v0.64.3: @senamakel, @M3gA-Mind.

Full Compare

v0.64.0...v0.64.3

The Intelligence Upgrade 🎉

v0.63.12 to v0.64.0 brings 674 PRs across 674 commits, with upgrades across memory, voice, agents, reliability, and developer foundations. Thank you to everyone who contributed to this release. ✨

Highlights 🚀

🎙️ Voice & hands-free control

Voice and hands-free control get more useful across the desktop, with always-on listening, global push-to-talk, faster command routing, and stronger automation paths for interacting with apps. (#5454, #5489, #5536, #5546, #5547, #5549, #5550, #5554, #5557, #5650, #5947, #5949, #5961, #5981, #6060, #6425, #6594, #6595, #6621) — Thank you @senamakel, @YellowSnnowmann, @JMak-Security, @M3gA-Mind, and @AdityaThakur193!

🧠 Memory & intelligence

Memory and intelligence flows are faster, clearer, and more resilient, from source sync controls and richer ingest to batched embeddings and delegated memory retrieval. (#5313, #5446, #5449, #5450, #5457, #5465, #5470, #5477, #5480, #5487, #5488, #5490, #5512, #5515, #5517, #5519, #5521, #5531, #5544, #5555, #5564, #5566, #5575, #5587, #5608, #5611, #5620, #5639, #5646, #5657, #5668, #5673, #5684, #5692, #5693, #5709, #5737, #5740, #5756, #5769, #5776, #5779, #5799, #5803, #5808, #5823, #5824, #5825, #5839, #5844, #5853, #5854, #5875, #5880, #5884, #5888, #5932, #5943, #5963, #5965, #5969, #5970, #5973, #5997, #6006, #6011, #6013, #6014, #6015, #6017, #6021, #6026, #6035, #6045, #6046, #6049, #6050, #6052, #6056, #6065, #6135, #6170, #6173, #6183, #6203, #6208, #6220, #6237, #6264, #6269, #6272, #6333, #6373, #6421, #6495, #6527, #6532, #6550, #6592, #6607) — Thank you @yh928, @senamakel, @YellowSnnowmann, @mysma-9403, @oxoxDev, @M3gA-Mind, @nocstah, @ntdatt812, @anisayakmitra-in, @ligjn, @CodeGhost21, and @sanil-23!

✅ Tasks, chat & agent workflows

Task, chat, and agent workflows become easier to steer and inspect, with better task boards, persistent run state, safer delegation, and clearer model/thread controls. (#2401, #4358, #5257, #5261, #5397, #5439, #5447, #5452, #5456, #5460, #5471, #5472, #5491, #5493, #5514, #5524, #5525, #5527, #5528, #5529, #5534, #5535, #5540, [#5542](#554...

The Smooth Release Bump Upgrade

v0.63.12 is here—a clean, streamlined release update to keep everything moving forward. 🎉

Highlights

Release automation & housekeeping ⚙️

This update is focused on release hygiene: bumping the version to v0.63.12 via the automated release workflow, with no feature or bugfix PRs included in this range. — Thank you @github-actions[bot]!

Contributor Credits

Thanks to @github-actions[bot] for shipping this release update.

Full Compare

v0.63.11...v0.63.12

OpenHuman v0.63.11

187 commits — 52 features, 90 fixes, 45 refactors, docs and CI. This release carries everything merged to main since v0.63.9, including the kernel restructure, ElevenLabs Voice Agents, and a large reliability pass across memory, inference and voice.

Note on the previous notes: an earlier version of this page listed only the two release-promotion PRs. That was an artefact of how release is updated, not a description of the release. The full contents are below.

Highlights

Voice

• ElevenLabs Voice Agents — core RPC, realtime session, and local-agent relay handler (#5407)
• Send the relay identity token where the backend can read it (#5428)
• Stop the Piper voice id leaking into non-Piper TTS providers (#5367)
• Repair the execute bit on extracted Piper binaries (#5216), handle U16 sample format in the fallback capture path (#4537)

Memory & embeddings

• Warn before the embedding budget silently stops memory (#5402)
• Stop a signed-in user's embeddings failing, and unstick the parked queue (#5427)
• Classify Ollama-unavailable embed failures actionably (#5398)
• Drain the coding-session backlog across bounded passes (#5368)
• Canonicalize memory identifiers symmetrically (#5275); stop paging on caller ingest-payload schema errors (#5280)

Inference & model routing

• Route Codex OAuth chat to the subscription backend (#5371)
• Error on a misconfigured vision model instead of substituting (#5262); route vision at vision-capable models (#5253)
• Explain background-role provider fallback and verify BYOK keys (#5254)
• Route local AI through TinyAgents RPC (#5240); local model reliability — /api/tags fallback and real Ollama model ids (#5219)

Flows

• Surface bounded loops from the tinyflows engine (#5366)
• Accept declared workflow inputs on every run path (#5304)
• Per-item fan-out — run an array of flow work concurrently (#5302)
• Canvas visual refresh — kind tiles, elevation, step numbers (#5303)

Agent & chat

• Send user image attachments as image blocks, not text (#5370)
• Preserve chat context across a Quick/Reasoning switch (#5401)
• Name a thread in three words (#5412)
• Name the chat-template rejection instead of blaming model/temperature (#5309)
• Filesystem-offload convention for worker artifacts on long-horizon tasks (#5217)

Kernel restructure

• Collapse 124 flat domains into 31 gate-aligned families (#5328) — pure moves, byte-identical wire surface
• Kernelize openhuman: cut the workflow dependency floor 418 → 280 crates (#5314)
• Realign DomainGroup with the family directories (#5332) and guard three unenforced consumers (#5364)
• Gate the macOS Contacts cohort behind contacts (#5365); host capability adapters and config seam (#5396)

Media, mascot & config

• Error instead of false success on the generation poll timeout (#5411)
• Accept PNG/GIF/JPEG/WebP custom avatars with file upload (#5420)
• Surface a user notice when a corrupt config.toml is reset (#5340); stop transient active_user.toml read faults from resetting the profile (#5337)

Channels & sessions

• Stop a missing edit route from looking like a deleted message (#5283)
• Keep the /auth/me 401 classifiable as session expiry rather than a hard logout (#5311)
• Stop out-of-credits background scouts paging Sentry (#5310)

Contributors

Thank you to everyone who shipped this release:

Full compare

v0.63.9...v0.63.11

The Flows, Memory, and Voice Power‑Up

This release is packed with deep workflow-building upgrades, a leaner kernelized core, and big reliability wins across memory, inference, and voice—plus a bunch of UI polish that makes everything feel snappier and sturdier. 🎉

Highlights

Flows & Workflow Copilot, leveled up 🧩

Workflow authoring and Copilot got a major usability and reliability pass: clearer step labels, a “Save & enable” path, safe live test-runs behind approvals, a Stop button, better run-rail behavior, improved compile/runtime correctness, and a visual refresh across the canvas. There’s also more robust store behavior (corrupt-row handling, transactional upserts), better localization coverage, and tighter run/approval boundaries so what you approve is what runs.

(#5088, #5089, #5090, #5097, #5099, #5111, #5112, #5114, #5115, #5120, #5128, #5135, #5148, #5150, #5153, #5176, #5205, #5212, #5223, #5227, #5228, #5248, #5264, #5265, #5285, #5286, #5287, #5289, #5290, #5292, #5293, #5294, #5296, #5297, #5302, #5303, #5304, #5366) — Thank you @graycyrus and @senamakel!

Memory & agent profiles that scale 🧠

Memory took a big step forward: TinyCortex consolidation and tool-memory cutovers continue, identifiers are normalized more reliably, and a new warning helps prevent “silent” memory stoppage when embedding budgets are exhausted. Flows also gained more intentional memory behaviors (flow namespaces + digest, dedicated memory node support, and read-only memory helpers), while profiles gained “hermes-style” homes that neatly organize per-profile workspace, memory, skills, and cron attribution.

(#5043, #5101, #5118, #5140, #5175, #5176, #5205, #5220, #5227, #5237, #5244, #5246, #5275, #5280, #5282, #5368, #5398, #5402, #5427) — Thank you @senamakel, @graycyrus, @YellowSnnowmann, @CodeGhost21, and @mysma-9403!

Voice, mascot, and realtime experiences 🎙️

Voice capture and TTS reliability improved (including better sample-format handling, Piper binary fixes, and cleaner provider behavior), and realtime voice mode is now surfaced by default. There’s also a big new integration: ElevenLabs Voice Agents with core RPC, realtime sessions, and a local relay handler—paired with UX improvements like sentence-by-sentence streaming to reduce perceived delay.

(#4537, #5216, #5367, #5404, #5407, #5428, #5441) — Thank you @mysma-9403 and @YellowSnnowmann!

Chat, conversations, and day-to-day UI polish 💬

Chat and conversation UX got smoother and sturdier: long URLs now wrap correctly, a render loop was fixed, template rejections are clearer, and the thread-goal footer no longer auto-scrolls unexpectedly. The “Human” experience also evolves: the mascot can now live docked inside chat, supports uploaded custom avatars, and avoids lipsync-related UI lockups—plus you can name a thread in three words.

(#5030, #5210, #5274, #5309, #5369, #5412, #5419, #5420) — Thank you @jgentes, @graycyrus, and @YellowSnnowmann!

Inference, embeddings, search & cost correctness ⚡

Inference and embeddings behavior is more predictable and safer: better model routing (including vision-capable models), more actionable errors (and redacted secrets), improved local model reliability, and clearer handling for BYOK keys and budgets. Search expands with Exa support (managed + BYOK), and attribution is improved so usage is correctly linked to resolved providers.

(#4885, #5103, #5117, #5206, #5207, #5215, #5218, #5219, #5222, #5253, #5254, #5262, #5363, #5371) — Thank you @M3gA-Mind, @YellowSnnowmann, and @samrusani!

Core kernelization, feature gates, and performance 🧱

The core continues getting leaner and more modular: big dependency trimming, kernelization work, DomainGroup realignment, and more functionality gated behind explicit compile-time features. Performance and orchestration also improved with batching, shared config arcs, checkpoint/resume for delegations, and runtime process pooling across agents and skill runs.

(#5029, #5040, #5052, #5061, #5062, #5068, #5078, #5079, #5129, #5178, #5202, #5233, #5240, #5314, #5328, #5332, #5364, #5365, #5396, #5405) — Thank you @senamakel, @YellowSnnowmann, @ox...