📝 docs: retry failed compact, oversized-turn stub, and send gate
📝 docs: Compact stays disabled until compaction finishes
📝 docs: compact token floor uses a message-id watermark
📝 docs: empty-at-ceiling on both paths and custom summarizer window
📝 docs: compact enqueue toast, token floor, and summarizer window
📝 docs: MiMo JSON-mode schema validation Note that ChatHub checks tool names and required arguments, so empty or invented JSON selections fail the group-chat supervisor turn.
📝 docs: MiMo JSON-mode supervisor and env-only Token Plan search
📝 docs: MiMo Token Plan search fallback and safe error.param logging Document ChatHub search on Token Plan, json_object structured output, and that generation-debug logs errorParamClass/hash only.
📝 docs: note Token Plan 400 from null frequency/presence_penalty
📝 docs: omit Token Plan native web_search and surface error.param Token Plan chat 400s come from web_search-without-plugin and temperature > 1.5, not only leaked fields. Use ChatHub/MCP search until Xiaomi Console enables Web Search.
📝 docs: troubleshoot MiMo Token Plan 400 Invalid request parameters
📝 docs: note vendored Xiaomi MiMo logo for ChatHub Settings
📝 docs: document MiMo web search plugin activation and thinking sampling Console plugin activation is required before ChatHub's native-search toggle, and thinking mode only drops temperature/top_p.
📝 docs: add Xiaomi MiMo provider setup and gear-panel notes
📝 docs: document dream settle fold and envelope fields CHATHUB_COMPACTION_DEBUG settle records now include keep-N counts, overflow envelope kind, and LLM-vs-concat fold path; already_has_card and regenerate reasons are allowlisted.
📝 docs: hide only checksummed opaque-v3 framing Leading v1/v2 and unvalidated v3-shaped lines stay ordinary range-card text.
📝 docs: treat leading [overflow:opaque-v1] as ordinary range-card text
📝 docs: keep [overflow:opaque-v1] payload lines as ordinary text
📝 docs: opaque overflow keeps tail even when a folded day looks like a heading
📝 docs: opaque overflow fold, sentinel uniqueness, and token-limit rewrite
📝 docs: overflow fold rewrites summaries over 6400 The memory dream asks the model to rewrite an over-long overflow summary instead of trimming it.
📝 docs: overflow range card re-summarizes to 6400 Keep N now folds older single-day cards into one rolling range summary when the dream runs, with a 6400-character cap and concat fallback.
📝 docs: keep stuffed [date:] overflow lines escaped in the editor
📝 docs: copy visible overflow text and keep [overflow:v1] as content
📝 docs: ordinary date-only body lines stay text even in multiples
📝 docs: note overflow [date:] headings in dream range cards
📝 docs: cap overflow card bodies at the per-card limit Note that the merged range card is editable because its body is capped the same way as single-day cards.
📝 docs: keep newest overflow and bound dream-memory headers Document oldest-first overflow truncation, the serialized (N+1) budget including card headers, and atomic Keep dream cards saves.
📝 docs: note bounded overflow truncation for dream-memory cards
📝 docs: dream-memory dated cards, keep-N, and per-card regenerate