📝 docs: skip calibration when skill text is unavailable A send still goes out when skill instructions were not already available; that reply is not used for the learned correction.
📝 docs: background sends honor the max-tokens switch A saved completion cap is not sent while the switch is off, including background generation.
📝 docs: describe stable token calibration refresh The badge reloads the correction when usage is reported, and a thread counts the topic messages that request will send.
📝 docs: note when auto-compact reserves max tokens Co-authored-by: Cursor <cursoragent@cursor.com>
📝 docs: describe input-budget auto-compact estimates Co-authored-by: Cursor <cursoragent@cursor.com>
📝 docs: scrolling keeps highlighted bubbles
📝 docs: stopping the wheel does not jump the thread
📝 docs: next scroll uses the current bubble height after resize
📝 docs: hover does not switch to light scroll markdown
📝 docs: light scroll markdown and PWA topic-sheet unmount Note that fast scroll keeps bubble chrome, and mobile keeps Back instant without leaving session rows subscribed.
📝 docs: planning clock survives topic switch; full tool bodies stay
📝 docs: tool results are sent in full
📝 docs: MCP tool results are not 8k-capped Builtin dumps still use the wire marker; MCP tools/call bodies stay full.
📝 docs: note token badge debounce on long topics The next-request chip waits ~300ms after typing so a long thread stays scrollable.
📝 docs: zero history and input-template changes recount the next-request estimate
📝 docs: durable send freezes token-estimate settings; history expand recounts older turns
📝 docs: token estimate stays bound to the request this tab sent (topic switch / reload fallback)
📝 docs: anchor stays bound to the sent request; correct the reload anecdote (1,000-token report vs >11,000 local estimate)
📝 docs: anchor re-verification after reload (fresh full estimate until session-observed report)
📝 docs: paged memory recall continuation, recovery summarizer window + RAG/summary retention
📝 docs: overflow recovery override + History Compress note, entry-scoped memory recall, strict MCP non-replay
📝 docs: trusted vs untrusted memory provenance
📝 docs: memory search/read recall, injection budget, dream dedupe
📝 docs: usage-anchored token estimate, overflow auto-retry, tool-result cap, summarizer deadline
📝 docs: dream falls back to recent-message excerpts; settle gains topicsWithExcerpt The memory dream no longer requires a topic compaction summary: active topics without one feed the dream through a bounded excerpt of their recent messages from that UTC day. no_summaries now means no readable content at all. The dream_scheduler_settled row gains topicsWithSummary / topicsWithExcerpt. Co-authored-by: Cursor <cursoragent@cursor.com>
📝 docs: keep token floor at cursor after last remaining delete Co-authored-by: Cursor <cursoragent@cursor.com>
📝 docs: say send is allowed below the configured high watermark
📝 docs: send-gate after compact, watcher latch, newest stub tail
📝 docs: retry failed compact, oversized-turn stub, and send gate
📝 docs: Compact stays disabled until compaction finishes