Project search now queries the full server-side session index instead of filtering whichever pages happen to be loaded. Dedicated watched queries keep search isolated from recents and pickers, while debouncing, pagination, polling, and cache keys all remain correct.
Sequence and slate editing now share a single per-agent conversation lifecycle with stable selection, create-on-first-message, creator-aware defaults, and consistent compact gutters. Deleted sequences remain safely inspectable through restore-only actions, while obsolete child-thread controls are gone.
Project context now stays coherent across surfaces: browser tabs identify the active session, and renaming an agent propagates to its row plus every live top-level sibling chat across clients while excluding subagent threads at the query boundary.
Auto-following became an explicit project preference rather than a fragile browser-local exception. Turning it off stops every automatic follow path and drops the current presence; users can still follow manually, and turning it back on resumes live-agent behavior.
Sequence editing now continues in each agent's durable conversation instead of spawning a new chat per sequence. The modal and action rail resolve the same thread, preserve legacy edit history, avoid closed-thread necroposting, and let InVideo behave like a normal renameable agent.
The rollout exposed an archival recovery branch that could never execute because agents are closed at the thread level, not archived at the row level. Removing the command and frontend push leaves fresh-chat creation to the real usable-thread rules.
Media mentions became a model-wide capability rather than a handful of image-model exceptions. Frontend labels, payload ordering, backend token resolution, pasted text, restored prompts, and Seedance frame slots now agree on the same Image/Video/Audio numbering contract.
feat(agents-models): extend @attachment prompt mentions to all models Previously only 8 hardcoded agents (nano-banana + gpt-image families) got the.
Chat attachments now retain the file identity needed to download original bytes. New entries use the workspace-scoped download path, older media falls back through a CORS-safe S3 route, and unsupported legacy cases fail visibly instead of silently opening a broken tab.
Voice-design outputs are now treated as previews that must pass through the clone-confirmation sequence before TTS. Moving the instruction into shared sub-agent serialization covers every agent version and correctly handles both MiniMax and ElevenLabs behavior.
Generation worker threads now capture the actual prompt when the first result lands instead of preserving an internal delegation ticket. Listing, Copy Prompt, and Reprompt therefore agree on the user's original intent across both generation-agent paths.
The generate-panel frame swap is now a real state transition: attachments exchange slots and their upload metadata is retargeted too, preventing synchronization from moving them back after the UI swap.
Raw BytePlus and MiniMax envelopes no longer leak into customer-facing errors. Known cases receive curated copy, unknown cases fail closed to safe generic messages, and the untouched provider payload remains available to engineers for debugging.
Support preview keeps the conversation readable in local context while exposing the exact UTC timestamp on demand, removing ambiguity during incident handoffs without adding visual noise.
The BI display regained near-real-time activity without reopening the replica-load problem. Bounded indexed polling fills only the gap after the warehouse watermark, Redis provides the freshness seam, and circuit breakers, timeouts, window caps, and kill switches contain failure modes.
Provider-reported input, output, cache-read, and cache-write tokens now travel with every LLM debug log and appear directly in the debug panel. Cache signals are clearly separated from InVideo's own cache-hit tag and survive copy-all export for diagnosis.
Usage badges now reconcile against the same thread-block boundaries as the charges panel. Backend steps expose a monotonic block total and the frontend takes the latest maximum per block, eliminating missing charges and repeated counting within a turn.
A full comparison bench now runs background-removal models against shared avatar clips without database state. It fans out resumable jobs, scores outputs with Gemini, tracks cost and duration, and gives reviewers synchronized side-by-side playback, scrubbing, retry, and rescoring.