The Reliability & Control Upgrade
This release is packed with sturdier workflows, smoother chat and Agent World experiences, and a ton of under-the-hood hardening to keep everything fast and dependable. 🎉
Highlights
Workflows you can trust 🧩
The Workflows copilot and canvas got a major reliability pass: proposals consistently render (even for larger graphs), every copilot turn now ends in a real terminal outcome (no more silent trail‑offs), Accept now saves automatically, and the runs UI stays accurate and live-updating. Safety and correctness also improved with fixes for live-arming on updates, better circuit-breaking, approval-state visibility, and stronger tier/approval enforcement for Composio tool calls.
(#4881, #4886, #4887, #4888, #4889, #4890, #4891, #4893, #4904, #4911, #4932, #4933, #4941, #4942, #4943, #4945, #4955, #4956, #4999, #5000, #5007, #5011, #5012, #5013) — Thank you @graycyrus and @senamakel!
Chat that feels snappier & more controllable 💬
Chat got a big UX win with a Stop button plus ESC-to-interrupt and re-edit, and stronger protection against “hydrate” snapshots wiping fresh streaming or tool timelines. Orchestration also became more transparent: async sub-agent failures and “awaiting user” outcomes now surface correctly in chat, and delegations are now async-by-default with durable, resumable sessions and persistent workflow proposals.
(#4898, #4903, #4908, #5074) — Thank you @M3gA-Mind, @graycyrus, and @senamakel!
Memory & transcript upgrades 🧠
Memory ingestion and recall got sharper and safer: OpenHuman can ingest Codex and Claude coding sessions, Composio sync no longer keys upserts off secret-like titles, and recall avoids pulling same-session “echo” documents that drown out real facts. On top of that, transcripts are now much more robust: cold-boot resumes restore full-fidelity context and thinking, and the transcript system is moving to an append-only, derived view with raw session logs as the source of truth.
(#4863, #4953, #4991, #5077, #5086) — Thank you @senamakel, @M3gA-Mind, and @graycyrus!
Agent World (Tiny Place) feels alive 🌍
Tiny Place is now more interactive and complete: DMs and feeds update in real time, ledgers and feeds paginate beyond the first 50 items, and directory entries finally open a rich profile view. Profiles also leveled up with editing for your own profile (name/bio/avatar) and a public viewer route for arbitrary handles.
(#4951, #4952, #4988, #4989, #4990, #4994, #4996, #4997) — Thank you @M3gA-Mind!
Performance, polish, and “it just works” fixes ⚡
From accessibility and UI polish to startup and runtime performance, this release reduces friction across the app: improved dark theme contrast, paused mesh-gradient rendering when the app isn’t visible, and boot polling is consolidated and quieter. First-run setup overlays now appear only when actually provisioning, channel setup guidance is corrected in-app, pricing/checkout redirects behave properly, and the settings “Clear App Data” reset now correctly targets the signed-in user.
(#4897, #4902, #4905, #4910, #4939, #4954, #5035, #5036, #5066, #5075, #5085) — Thank you @M3gA-Mind, @YellowSnnowmann, and @senamakel!
Build flexibility, CI hardening, and new surfaces 🛠️
The core is now far more modular and easier to ship in slim builds via compile-time feature gates (media, web3, flows, skills, MCP, meet) and a new feature-gated terminal chat UI (openhuman tui). Desktop shipping and CI also got stricter and safer: voice and tokenjuice-treesitter are correctly forwarded into the desktop build, CI now gates feature-forwarding drift, macOS clippy blockers are resolved, the “gates-off” smoke lane is repaired and runs tests, and learning startup subscribers are decoupled from channels and covered with more stable tests. There’s also a core refactor moving the web chat conduit out of channels/.
(#4840, #4855, #4912, #4913, #4914, #4915, #4916, #4917, #4934, #5004, #5020, #5023, #5025, #5070, #5082, #5084) — Thank you @oxoxDev, @CodeGhost21, @senamakel, and @M3gA-Mind!
Integrations, inference, and tooling improvements 🔌
Integrations and local AI got more reliable: OpenAI-compatible local runtimes now probe /v1/models correctly, custom embedding endpoint verification errors are clearer and actionable, and streaming retries now honor Retry-After. Composio tool execution is smarter and safer with per-action tools gated on their full contracts (reducing argument guessing), and unsyncable toolkits no longer auto-register as memory sources. There’s also a new generate_document tool that emits real .docx, plus a public accessor for last-turn usage totals for downstream hosting/billing.
(#4899, #4909, #4940, #4993, #4995, #5064, #5067) — Thank you @M3gA-Mind and @senamakel!
Account, sharing, and navigation delights 🔗
You can now generate one-tap “look what my agent did” share cards for X and LinkedIn, the X/Twitter “Continue with Google” popup stays in-app, Rewards is finally discoverable via a cloud-gated sidebar entry, and the Codex auth button is properly localized. Plus: proactive “Agentic task insights” fixtures and lifecycle resets were tightened to keep the UI consistent.
(#4892, #4944, #4949, #4992, #5010, #5015, #5065) — Thank you @YellowSnnowmann, @graycyrus, and @M3gA-Mind!
Safety & prompt discipline changes 🧯
Desktop automation safety controls saw churn: an Emergency Stop was introduced and later fully removed based on product direction. Meanwhile, agent behavior improved for local LLM providers with provider-aware context/delegation guardrails, stronger structured-output enforcement, more resilient tool-call parsing for small models, and AGENTS.md instructions are now loaded into the system prompt—with follow-up security hardening to prevent symlink/path escape and to redact debug logs when AGENTS.md is present.
(#4600, #4900, #4935, #4938, #5073, #5087, #5096) — Thank you @CodeGhost21, @M3gA-Mind, @sanil-23, @yh928, and @senamakel!
Additional highlights 🔗
A few more focused contributions round out this release with targeted fixes and improvements that are worth calling out. (#4872, #4877, #4906, #5008) — Thank you @senamakel, @graycyrus, and @M3gA-Mind!
New Contributors
- Thank you @yh928 for improving tool-call robustness by recovering pipe-garbled
<tool_call>tags from small models. (#4938)
Hope you’ll join us on Discord—there are exclusive roles and contributor rewards waiting.
Contributor Credits
Thank you @CodeGhost21, @cyrus Gray, @mega Mind, @oxoxDev, @sanil-23, @steven Enamakel, @YellowSnnowmann, and @yh928!
