📝 docs: document Codex 5-hour and weekly remaining on Settings
📝 docs: distinguish Codex stream debug outcomes Record that chat_stream_settled follows the Responses lifecycle, not SSE event count, so operators can tell failed streams from empty and ok.
📝 docs: document CHATHUB_OPENAI_CODEX_DEBUG for empty Codex chats Platform OpenAI stream flags do not cover ChatGPT subscription traffic; enable the new safe switch and inspect overlay plus SSE counts.
📝 docs: recover Codex unique-index 42P07 startup loops by upgrading If ensureOpenAICodexOAuthTokens fails because the unique index already exists, upgrade the image; do not drop the table or edit the migration journal.
📝 docs: recover a stuck ChatGPT refresh by signing in again An interrupted server refresh can leave Settings connected while chat keeps failing; retries and another restart do not clear it, so operators must sign out and sign in on Settings → OpenAI.
📝 docs: distinguish temporary Codex refresh outcomes Temporary token-endpoint failures keep the subscription and continue on Codex while the access token is valid; only an already-expired access token fails that chat.
📝 docs: require ChatGPT device-code permission for Codex login Document the official ChatGPT device-login prerequisite and keep structured supervisor calls on the Platform key.
📝 docs: add ChatGPT Codex subscription login Document the Settings → OpenAI device-code flow, quota vs API key, and chat-only limits.
📝 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: describe planning-next-step status and the tool-round limit
📝 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: MCP model-boundary hygiene (description scrub, _meta strip, stable tool order)
📝 docs: usage-anchored token estimate, overflow auto-retry, tool-result cap, summarizer deadline
📝 docs: MCP circuit breaker, cache hygiene env vars, cancellation