Read conversation activity
The conversation keeps the agent's explanations and results in the main timeline. Routine terminal input, waits and completion events live inside the existing action blocks. Expand a block to inspect its calls and lifecycle details in order.
Follow live work
The latest action block follows work while the response is running. A finished block folds into its summary. You can expand or collapse it yourself; your choice takes precedence over automatic folding. Terminal waits still update the response's live status. A completed wait becomes a detail inside the action block rather than another standalone line in the conversation.
Failures, warnings and notices offering a recovery action remain visible in the main timeline. Closing a terminal through its controls records a stopped outcome instead of a process failure.
Inspect the sequence
Adjacent equal system messages share a count. Their timestamps do not determine equality, and a
different message interrupts the sequence: A → A → B → A becomes A ×2 → B → A. Different
statuses and recovery actions remain distinct. Unique secondary details are retained in the notice
tooltip.
Within explicitly linked assistant continuation segments by the same author, a repeated item ID belongs to the segment that first recorded it. The history uses its latest outcome once, keeping the new segment's new work in order. Independent messages, different authors, and calls with the same wording remain separate.
Find and recover earlier messages
Sidebar search filters titles immediately and searches stored message text after you enter at least two characters. It applies to the open history tab; the External tab filters native CLI sessions by title only. Found in messages shows one row per conversation from up to 30 ranked message hits; many matches in one conversation can leave fewer result rows. Selecting a row opens the conversation, not the exact matching message. Rows whose sessions are absent from the loaded sidebar list are omitted. Search uses quoted full-text terms together, not a substring, regular expression, or semantic similarity search. Punctuation-only input has no searchable terms. An unavailable index or a search failure can appear as no results.
An agent with GlobalSearch can recover earlier text when you refer to past work, follow a
checkpoint pointer, or inspect another session in your workspace:
| Request | Result | Boundary |
|---|---|---|
query | Ranked sessions, short snippets, and nearby text | Defaults to five sessions; discovery is capped at ten |
around_message_id | Messages before and after that message | The session is inferred; an explicit session must match |
session_id alone | A window ending at that session's latest message | Defaults to five messages on each side; maximum fifteen |
| No arguments | Recent sessions with previews | Defaults to ten; maximum twenty |
These reads are scoped to the Localix user. They query stored history independently of the current model view: a checkpoint or topic change does not erase the older rows. Archived sessions and inactive message branches can still contribute results. A window follows stored message sequence, so it can cross branches or topic boundaries rather than reconstructing one active path. Discovery examines a bounded set of ranked hits; it is not an exhaustive inventory.
The search index covers message text, not tool inputs/results or attachment and context-binding
bodies. Agent previews are clipped to 200 characters in discovery and 500 in scroll mode, with an
8,192-character total JSON budget. Large discovery replies drop bookends, then nearby messages, then
trailing sessions with an omission notice. An oversized single result returns a truncation notice.
Use narrower queries or ToolOutput for an original tool artifact; these previews are not an exact
replay of a complete turn.
Model context
Client-only terminal lifecycle notices are presentation metadata and are not sent to the model. Tool calls and their results still reach the agent's history. Repairing a repeated continuation prefix prevents those same calls and results from appearing twice in the model context as well. Independent terminal observations keep their normal checkpoint and output-size policies.
See Context checkpoints, Sessions and projects, Background work and Interactive terminals.
Session recap
When you come back to a chat after a break, a one-line recap of where you left off (at most 400 characters) can be waiting in the transcript, marked Recap. It is written in the background once the window has been unfocused for three minutes after the last finished turn. It needs a session with at least three turns, is never written twice in a row, and never in a run that nobody watches (a loop, a scheduled run, an API call).
A recap is a record for you: it is never sent to a model and never enters the conversation's context. It uses the Session recap route in the model picker, which has its own model, effort and an on/off switch on that row; with no model chosen it uses the session's model. The switch is on by default.
/recap writes one on demand, even when automatic recaps are off. It answers with a notice when
there is nothing to sum up yet or no model could be reached.