v2026.10.5

@code-yeongyu/senpi

Breaking Changes

Added

• open_session accepts browserEngine (connected, builtin or none) and hosts advertise the browser_engine capability: the choice is per session, so that session's tool subprocesses and eval kernels see OMO_BROWSER_ENGINE and no other session does, a reattach moves or keeps it, and BSK_HOME / BSK_BIN pass through unchanged (#2611).
• A media_placeholders client now gets an image_ref placeholder whose path names the stored image (written before the placeholder is emitted under the session's media/<durableSessionId>/ directory, private and read-only, deleted with the session), or an unavailableReason of image_too_large, session_limit or storage_error, so tool images other than read can be rendered from disk.
• An edited assistant response can be continued with no new prompt: the RPC continue_from_leaf command (advertised as the continue_from_leaf capability) starts a turn from the session leaf through a hidden nudge the transcript never shows, refusing with streaming, nothing_to_continue, or leaf_not_assistant (the conversation ends on a user message). get_available_models rows report supportsAssistantPrefill (false for every model today) (#1930).

Changed

Fixed

• nvidia starts on a model its catalog still has: the v2026.10.4 catalog no longer lists nvidia/nemotron-3-super-120b-a12b, so an NVIDIA-only user without a saved model silently started on the catalog's first entry (deepseek-ai/deepseek-v4.1-flash) instead of the provider default. The default is now nvidia/nemotron-3-ultra-550b-a55b, and a release now stops when a catalog regeneration drops any bundled provider's default (#2645).
• apply_patch keeps each line's ending when it updates a file: CRLF and mixed-ending files are no longer rewritten to LF, inserted lines take the file's line ending, and context lines stay exactly as they were, so a one-line change is a one-line diff (#2638).
• apply_patch no longer drops a file section whose header is indented while reporting success: indented headers apply as in Codex, any other line between file sections rejects the patch before anything changes, and the permission system's per-file approval lists exactly the files the patch writes (#2636).

Removed

@code-yeongyu/senpi-ai

Breaking Changes

Added

• Models can declare supportsAssistantPrefill, and modelSupportsAssistantPrefill(model, { thinkingEnabled }) reports whether a request may end with an assistant message the model continues (never with extended thinking on the Anthropic Messages API); no built-in model declares it yet (#1930).

Changed

Fixed

• A first-turn forced tool_choice that a strict-schema gateway refuses by naming the forced tool (its tools.N index or quoted name in a 400) is retried once without the forced choice, like the other forced-choice refusals, so a new session's first message no longer fails there; an unrelated 400 still fails (#2648, reported by @rhyme227 in code-yeongyu/oh-my-openagent#9507).

Removed

@code-yeongyu/senpi-agent-core

Breaking Changes

Added

Changed

Fixed

Removed

@code-yeongyu/senpi-codemode

Breaking Changes

Added

• Eval kernels expose the session's OMO_BROWSER_ENGINE and clear a value inherited from the host process for a session that chose no engine (#2611).
• A Python, Ruby, or Julia eval kernel whose interpreter dies is replaced once instead of failing every later cell: Ruby and Julia no longer stay closed after a crash, and a Python kernel whose stuck interpreter finally exits recovers instead of rejecting every cell. Cells queued behind the death keep their order and run on the replacement, whose first result says [<language> kernel was restarted after <reason>; every global is lost]; the cell that was running fails once and is never re-run, and a replacement that dies before finishing a cell fails the queued cells with eval_kernel_unavailable (#2452)

Changed

Fixed

Removed

@code-yeongyu/senpi-tui

Breaking Changes

Added

Changed

Fixed

Removed