You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Comment on documents as you read. In a Markdown or HTML artifact, tap a
paragraph or select words to write or dictate a note. In a PDF or image, Mark an area to comment lets you drag a box; the note sends a crop and,
for PDFs, any text found inside it. Keep drafts on your device, review them in the note
tray, then send them together to the conversation that delivered the file,
or use Send now for one note without leaving the reader. Numbered
questions in Markdown and HTML offer Yes, No and Your call quick
replies. See Commenting while you read.
GPT-6.1 Terra, available as terra or gpt-6.1-terra, with Fast mode.
GPT-5.6 Terra remains selectable by its ID.
Claude Fable 5.2, available as fable or claude-fable-5-2, with Fast
mode. Fable 5.1 remains selectable by its ID.
Meta Muse Ember 1.0, available by ID as muse-ember-1.0; muse still selects Muse Spark.
MCP tools load on demand on supported Anthropic models and the ChatGPT plan. The agent
uses tool_search to find tools by keywords, server or exact name, instead
of putting every tool definition into its context up front. Other providers
keep the full tool list. See MCP tool search.
More detail in CLI JSON-lines output: subagents' tool calls carry their
child ID, and compaction reports the summarizer and its token usage, with
job IDs to link background summaries to their usage events.
Changed
Repeated tool calls are checked with their results. The agent gets a
warning after the third identical call-and-result batch and stops after the
sixth, whether the result is a success or an error. Changing results do not
count as a stuck loop; status/wait tools and batches you fully refused are
exempt. See Repeated tool calls.
For you context is saved as snapshots. Changes to a session's open
questions and recent News are appended to its history rather than rewriting
an earlier request. These internal snapshots stay out of the visible chat
and are refreshed after compaction.
Agent instructions are aligned across owners, sessions and subagents:
human-only approvals still require your own answer, an open question does
not authorize consequential work, and memory is not a substitute for
canonical project sources. Redundant tool and style guidance is removed.
Mirada returns in place of Serena as an owner's animated face.
Security
Updated DOMPurify to 3.4.17 and golang.org/x/net to v0.62.0.
Fixed
More stable prompt-cache prefixes. Unchanged context inventories stay
anchored across requests and runs. Restored sessions reuse their own saved
MCP tool definitions while servers reconnect, withdraw them if a server
fails, and bound the wait for a connection rather than waiting indefinitely.
Anthropic API-backup keep-alive can keep the paid cache warm when
keep-alive is enabled and the backup is admitted; it stops pinging a key
observed to be out of credit. Results from an older sign-in/key pair no
longer overwrite the current pair's status.
Anthropic response messages retain the plan-quota utilization and reset
times reported with that response. Detailed cache diagnostics are off by
default and can be enabled with MOA_DEBUG_CACHE=1.
Subagent model choices survive switching between All models and Selected only on the same device.
An owner's session tally includes idle conversations with background jobs
still running, without counting those jobs twice as live updates arrive.
Restored provider continuations resume without inserting fresh For you
context into a request that must be resubmitted unchanged.