• Fixed model searches ranking stronger matches ahead of weaker signed-in matches while preferring signed-in providers for equivalent results (#539 by @eliebak).
• Fixed large IPython variables repeatedly slowing later turns by excluding them from persistent snapshots and removing them when context is compacted.
• Fixed daemon socket paths being used verbatim in identity derivations: on supported platforms, --daemon-socket spellings differing only by duplicate or trailing slashes now normalize to one canonical path, so worker-descriptor namespaces, daemon log files, and persisted descriptors agree.
• Added a thinking option to rlm.run for spawning subagents with an explicit reasoning level; invalid levels for the resolved child model fail spawn.
• Changed opening the agents view (full or scoped) with a draft prompt to auto-stash the draft instead of refusing; the draft is restored into the editor when the session is reopened.
• Fixed Shift+Enter no longer inserting a newline in terminals that send a literal \n (for example a Ghostty shift+enter=text:\n mapping): the byte decoded as ctrl+j and triggered the new edit-diff toggle instead of the editor newline.
• Removed a system prompt paragraph referring to an async bash() kernel helper and managed jobs that do not exist in the runtime.
• Changed RLM guidance to orchestrate independent workers in parallel, use available async shell helpers safely, end the turn instead of sleeping, polling, or blocking on long awaits, provide proactive outcome-focused progress updates from root agents, and use simplified technical English for user-facing prose.
• Fixed new top-level daemon sessions inheriting an RLM child depth from the supervisor process.
• Fixed active goals stalling after a mid-goal automatic compaction when the previous continuation prompt was already running: only undelivered continuations deduplicate, so a fresh continuation is queued instead of being suppressed.