Messages and feedback

Give feedback while an agent works. The message can join its next model iteration; it does not have to wait for the whole answer to your original request. An authorized agent can use the same durable mailbox to coordinate with another session.

Busy, idle and paused recipients

Recipient	Message delivery
Busy	Queued before its next model call at a safe iteration boundary. An active foreground tool call must return first.
Idle	The queued message can wake the recipient automatically.
Paused	The message remains queued. Sending ordinary peer mail does not remove the hold.
Terminal subagent	Addressing the child can create a linked successor while preserving the earlier run and result.

A durable acknowledgement means CLIO accepted the message. It does not mean the recipient has read or acted on it. Messages carry source and target identities; replies can refer to the original message and correlation ID.

Find an earlier conversation, then address it exactly

Find the session where we investigated the calibration mismatch. Read its summary
and ask that session to check whether its conclusions apply to this dataset.
Use the exact session returned by the search, and report the reply here.

This is an example request. Discovery and messaging require the corresponding tools among the conversation’s plugins and the permitted session scope.

You can also ask without knowing the session’s name:

The other day we talked about catering costs and contingency. Find the agents
that worked on it, read what they decided, and show me the relevant note.

The model chooses topic words for the search, then checks the returned transcripts. It should identify matching sessions with dates and source passages; it should not treat the first match as proof of the remembered decision.

Discover readable sessions with query_tasks(kind="Session", scope="authorized", query="relevant words").
Read the returned handle with observe_tasks(view="transcript") or observe_tasks(view="summary"). Use view="artifacts" for linked deliverables.
Address the exact authorized handle or stable session with send_agent_message.
Continue other work. The recipient can reply through the mailbox; an idle sender can wake to receive that reply.

Search a discussion within a date range

scope="authorized" searches across the sessions you can read, not just the current conversation. kind="Session" limits results to root assignments; include Subagents by omitting kind when searching message dates.

For example, the agent can narrow the topic search to October 19 in Chicago:

query_tasks(
    scope="authorized",
    query="catering contingency",
    message_after="2026-10-19T00:00:00-05:00",
    message_before="2026-10-20T00:00:00-05:00",
)

Both timestamps are optional. The start is inclusive and the end is exclusive; each must include a timezone. The filters select when retained messages were written, so a recent discussion inside an older session can match. With a date filter, topic words must occur in the selected messages; an old description or title cannot make an unrelated recent discussion match. A date-only query finds readable Session/Subagent assignments with messages in the window.

Results still identify finite assignments. Read their transcripts to verify the passages and group multiple assignments belonging to the same conversation. They retain assignment creation order, rather than relevance or discussion-date order. Follow each cursor with the same query and time bounds; changing them is an error. An empty or reversed date window is also an error. Time filters narrow lexical search; they do not add semantic similarity retrieval or grant access to another session.

Current task reads and retained compatibility names

Query tasks searches readable descriptions and retained transcript text and returns handles usable by the shared controls. Search is lexical: all supplied search terms must occur in an assignment’s searchable text. It is not semantic similarity retrieval. The agent can broaden a query or try alternate wording. Optional message_after and message_before narrow the retained conversation messages being searched. Omit them to retain description and title matching.
Observe transcript, summary or artifacts reads original typed transcript rows, recorded summaries/outcomes with coverage, or linked deliverables. Follow returned cursors; a recorded summary is not a new whole-session summary.
The old session_search, session_read_summary and session_read_context_frame names remain explicitly selectable compatibility interfaces, not default tools. The old summary still means recent activity; the context-frame reader still inspects assembly provenance rather than all referenced contents. Their presence does not establish long-term memory.

Use exact source references for deeper reading. A short summary is a starting point, not proof that every earlier decision is represented.

Scope and ownership

Knowing a session ID does not grant access to it. Peer messaging checks the recipient’s human-recorded coordination rights. Model-supplied intent grants no access. Reading another session does not grant permission to pause, restart or cancel it. Recipient execution retains its own permissions; a peer message is not new human authority.

Task observation and waits default to the caller’s own task tree. Explicit, authorized broader reads can discover finite Session run handles, observe or wait on them, and subscribe independently to their results. A stable session address is a mailbox; a run handle identifies one execution and its outcome.

Retention and backpressure

Core retains message bodies, assignment history and delivery records in the existing session hierarchy. Delivered messages do not count against pending message capacity and are not deleted to make room for new messages.

gact.mailbox.max_pending_messages=0, or leaving it unset, allows an unlimited pending mailbox. A positive value limits undelivered peer messages, including messages staged for an iteration whose input has not committed. A full mailbox returns retryable backpressure for new peer messages. Accepted task results, human feedback and human controls continue to be admitted.

Full native-replacement qualification remains incomplete. Recorded root and child demonstrations cover messaging, observation-only privacy, revocation and independent completion delivery. A later two-workflow run also passed busy-peer delivery and historical-handle reads. Provider, recovery and current lifecycle/UI gates remain. See Session coordination.