Changed
ulw-plan now plans for the person the result touches, and the plan reviewer checks that it gets them there. (#8773) Before asking you anything, the planner names who the change affects - a customer, another programmer, a program or agent consuming the output - how they use it today, the state in which nothing snags, regresses or degrades for them, and every gap from today, each with its reason. Forks are resolved against that state first, so you are asked only the decisions that are genuinely yours; "MVP" or "phase 1" cuts are never invented, and when the ideal state is larger than what you literally asked for the brief says so in one line and plans it anyway. The plan file carries the user and the ideal-state rows under ## Scope, a ## Success criteria table that maps every row to the task that delivers it and the QA scenario that proves it (that section used to be empty), and a final Ideal-state fidelity gate whose shortfall becomes new tasks, never a note. The plan reviewer no longer approves at "80% clear": it checks who the end user is, what changes for them, which problem is solved, that every ideal-state row is delivered and proven, and that the approach can actually reach that state - a wrong approach for the stated user is a blocker, while a different approach that would also work is not its concern. The same rules ship in the OpenCode and Codex editions of the skill.
With no model profile set, OmO starts on the recommended model. (#8770) A fresh OmO Desktop or headless session with no model_profile now runs Recommended instead of Daily · Normal: Claude Opus 5.5 (medium), then Claude Fable 5.1 (xhigh), Kimi K3 (max), GPT-6 Astra (xhigh), GPT-6 Sol (medium), and GLM 5.3 (max), whichever you have connected first. The Claude subscription wins over the Anthropic API, and gateway aggregators such as OpenGateway and OpenRouter are never picked for this default. The OmO Native engine moves to senpi 2026.9.24, so the terminal follows the same order through the engine's recommended-model switch (senpi#2074). The four Daily/Geeky lanes are unchanged and apply when you pick one; model_profile: "recommended" selects the default explicitly.
The OmO Native footer badge credits its author and tells a dev build apart. (#8771) The status row now reads (😺 OmO Native by Q Kim), and a session started from an omob dev build shows (☕ OmO Native by Q Kim) instead, so you can see at a glance which binary you are running.
Fixed
Memory writes no longer fail with "'description' exceeds 1024 characters" when the model garbles the call. (#8774) Most ○ Not remembered notices came from one model mistake: closing the memory tool's description argument with </description> instead of </parameter>, which folds the note's whole body into the description. The memory tool now splits such a call back into its description and body and saves it, and tells the model what it repaired. A call that cannot be repaired unambiguously is refused as a garbled call rather than a too-long description, so the model resends it instead of trimming, and the notice says so in a plain sentence (The description was 6,242 characters; the limit is 1,024.) instead of the raw engine error.
- 549a60b Merge pull request #8783 from code-yeongyu/release/v5.0.0-beta.89-source-state
- 2d03f2c Merge pull request #8780 from code-yeongyu/feat/ulw-plan-affected-user-ideal-state
- de579ac feat(ulw-plan): anchor every plan on the affected user's ideal state (#8773)
- 6f622d6 Merge pull request #8778 from code-yeongyu/fix/memory-leaked-description-args
- f668fec fix(memory): repair arguments that leaked into description and name scaffolding before length
- 914ae54 Merge pull request #8776 from code-yeongyu/feat/native-badge-omob-coffee
- 7b4a44b Merge pull request #8775 from code-yeongyu/feat/recommended-default-profile
- 412e7fd Merge pull request #8761 from code-yeongyu/docs/readme-opengateway-sponsor
- f8e492a docs(readme): reword OpenGateway as sponsor and drop LazyCodex banner
npm i -g omo-ai@beta