📝 docs: explain Claude earlier-reasoning drops and readable background errors Fable 5.1 and the 5.5 models drop mismatched earlier thinking instead of failing; background-turn errors now show the provider message.
📝 docs: point Sandbox setup at OpenSandbox only The Dify sidecar is no longer a backend. Operator pages now use the Sandbox tool and OPENSANDBOX_* settings.
📝 docs: describe the longer tool-round budget and closing reply Co-authored-by: Cursor <cursoragent@cursor.com>
📝 docs: overlapping regenerate must keep the latest turn
📝 docs: click footer time and tokens for usage panel
📝 docs: note live-tail bootstrap resyncs finished replies
📝 docs: note reply footer timings
📝 docs: note SSE live-tail after recreate
📝 docs: correlate first-send terminal by operationHash
📝 docs: note first-send attach survives topic remount
📝 docs: note first new-topic send shows assistant immediately
📝 docs: planning clock survives topic switch; full tool bodies stay
📝 docs: describe planning-next-step status and the tool-round limit
📝 docs: leave session keeps browser-fallback tool continue Clarify that switching chat/session in the same tab is not Stop for browser-fallback turns (Kagi, DALL·E chat tools, non-HTTP MCP).
📝 docs: empty compaction summaries fail fast instead of eight retries Document History Compress as the summarizer model, thinking-model empty summaries, and that background compaction does not retry empty content the way title-transcript races do.
📝 docs: Stop writes the cancelled placeholder immediately Record that Stop does not wait for conversation cancel, and that stopped diagnostics apply only to the cancelled placeholder.
📝 docs: Stop records locally then writes a cancelled placeholder
📝 docs: Stop still blocks unpaid creates if the save fails Reload after Stop does not auto-start that image even when the conversation save is interrupted. A later generation that crashed without Stop still resumes.
📝 docs: Stop does not auto-restart unpaid chat images Remount and reload after Stop wait for Retry; running tasks still resume, and recovery creates join on the persisted send span.
📝 docs: diagnose chat Image tiles with generation-debug Point operators at CHATHUB_GENERATION_DEBUG chat_image_* events so leave-topic persist and Safari Load failed stay distinguishable.
📝 docs: chat Image tool keeps generating after leave-topic
📝 docs: sandbox provider, portal file scope, and paging Document SANDBOX_PROVIDER=dify, main vs portal Code Interpreter files, and newest-first paging on long topics.
📝 docs: add Code Interpreter DifySandbox sidecar Document the Compose sibling, operator setup, and that Graphile runs Python on the sandbox instead of the tab.
📝 docs: Code Interpreter no longer forces the whole reply onto the tab Search-only turns stay on the server when CI is only enabled; leftover Safari Load failed streams are interrupts, not unknown API errors.
📝 docs: RAG retrieve stays loading after you leave a topic Document that browser-fallback Knowledge retrieval keeps the assistant loading after a topic switch, and that tapping the mobile topic subtitle no longer shows a floating black title overlay.
📝 docs: generation-debug covers off-screen RAG, search, and persist skip
📝 docs: builtin tools and Knowledge Base also continue after you leave a topic With the tab open, KB retrieval and built-in tools keep going in the background; you do not need to switch back.
📝 docs: model continues immediately after MCP when you leave a topic With the tab still open, the summary starts after the tool returns; resume on return is only a backup.
📝 docs: log when MCP tools finish but the model never summarizes CHATHUB_GENERATION_DEBUG now records tool_loop_continue and skip reasons for the leave/return stall after a successful Tavily (or other MCP) call.
📝 docs: browser-fallback tools now continue the model after you leave