Hotfix: reopening a session with an unanswerable question no longer crashes the TUI. (#9268) When a session was resumed with a pending question whose saved arguments no longer formed a valid question set, the question widget showed "0 unanswered" and crashed with Cannot read properties of undefined (reading 'question') as soon as you expanded it. The call now settles as lost in a restart, so the model learns the question is gone and can ask again, and clicking a widget whose questions all have answers submits them. Thanks to @copycatcode for the report. This release runs on the senpi 2026.9.30 engine.
Added
A new accept-edits permission preset lets the agent read and edit files in the project without asking, while shell commands, paths outside the project and other tools still wait for your approval. (senpi#2430)
Chat bridges can ask the engine for a chat prompt surface, which drops the routing line, the handoff block and todo lines from replies meant for people in a conversation. Terminal and app prompts stay as they are. (senpi#2398)
Changed
The engine's recommended OpenAI model is GPT-6.1 Sol at medium, one slot below GPT-6 Astra. Models you listed yourself in recommendedModels stay as you set them. (senpi#2390)
ultrawork reuses QA and review evidence per target instead of rerunning everything after each patch, while keeping the nets that catch what the changer cannot see: each artifact records the commit and what it exercised; after an increment the session reruns the tests of every touched file and its importers, the scenarios that exercise them, and anything whose dependencies or environment moved, cites the capture for the rest, and still runs the full set once before the final message. Every re-review spawns a new reviewer with the delta diff and the cited blockers, at most twice. Defects inside the change's blast radius are fixed in the same run to the ideal state; defects outside it get a tracked issue and a line in the final message instead of growing the run. The memory line now also records every regression a check caught and each QA scenario with its invocation. (#9294, #9298)
Fixed
On a Claude subscription, a write or edit to a file outside the working directory runs once and returns one result. Before, Claude Code's own read check refused the call while the change still went through, so a retry could apply an edit twice. Thanks to @haamsuk-collab. (senpi#2401)
A running session keeps working through an update that changes how the engine's dependencies are laid out. Bash, monitor and eval used to fail with ENOENT until you restarted. The first launch after an update takes 2 to 3 seconds longer once while the engine copies its dependencies. (senpi#2408, senpi#2409)
An answer sent to a terminal session's question through its control endpoint reaches the model with its text, and answering such a question no longer triggers the first-turn todo reminder. (senpi#2407, senpi#2419)
The prompt-cache keep-alive pings in real sessions again, with the same tool list as the turn it keeps warm. Thanks to @MoerAI. (senpi#2389)
An eval cell's return value reaches the model whole up to the normal output budget instead of stopping after 768 bytes, and any output that is still cut says so. (senpi#2402)
On macOS, computer use keeps its Accessibility and Screen Recording permissions across updates, because the signed desktop engine now always runs from the same path. (#9282, #9288)
On Windows, killing a running background task marks it as killed instead of reporting a crash. Thanks to @Dante-dan. (#9228, #9233)
- c5d4972 Merge pull request #9305 from code-yeongyu/release/v5.1.6-source-state
- 7750a35 Merge pull request #9298 from code-yeongyu/fix/ultrawork-evidence-reuse-blast-radius
- dffb638 fix(ultrawork): fold the per-criterion test rerun into the per-target evidence rule
- 782d950 docs: note the ultrawork evidence-reuse, fresh-reviewer and blast-radius scope change (#9294)
- 7d293bf fix(ultrawork): reuse evidence per target, spawn a new reviewer per round, scope defects to the blast radius
- aa36eb4 Merge pull request #9288 from code-yeongyu/fix/stable-desktop-engine-20260930
- 53752e2 test(desktop-engine): make native transaction fixtures portable
- cc5175a fix(computer-use): launch signed release engines at a stable path
- c643633 Merge pull request #9233 from Dante-dan/fix/9228-windows-rpc-kill
- 99e3dde Merge remote-tracking branch 'origin/dev' into HEAD
- 8550c21 Merge origin/dev into fix/9228-windows-rpc-kill
- ffb925d Merge remote-tracking branch 'origin/dev' into fix/9228-windows-rpc-kill
Thank you to 1 community contributor:
- @Dante-dan:
- Merge remote-tracking branch 'origin/dev' into fix/9228-windows-rpc-kill
bun add -g omo-ai