The navigator's Agents tab reads the Claude Code, Codex and Copilot Agent Chat transcripts on this machine. Active now keeps working and waiting agents at the top, followed by Recent Agent Chats. Within these groups, cards keep their reading order during updates; activity timestamps re-sort them when a chat starts or stops running, or you change the project scope or period. New idle chats are appended without moving existing cards. Agent Chat names appear when the client stores one locally; summaries are labelled. Verbose shows document/task names and the ending preview immediately; it does not load full answers for every expanded card. Compact shows the same full cards for working/waiting and newly idle Agent Chats; older cards start collapsed. The card heading's single chevron expands or collapses both names and the answer together. Click an object's name to open it. Each name shows its status; hover over or focus it for the rest of what the Docs and Tasks lists show: type, priority, checklist progress, comments, areas or tags, and the audit flag. An ID expands a collapsed card; on an expanded card it copies the ID and briefly confirms the copy beside it. The agent-icon History link opens Agent History, newest first, with the full object title prominently shown. Search Agent Chat and object names, answer excerpts, agent, project, status or T/D ID. Expanded cards show the ending of the latest answer as Markdown in a compact preview without an internal scrollbar. Earlier text clips upward so the final question or next step stays visible. View beginning reads the full answer locally on click and opens it from the start with a scrollbar. View end returns to the ending preview. Your reading position in full view stays in place during live updates. An internal 256 KiB limit protects against unusually large answers: both the start and end remain readable, with an explicit notice where the middle was omitted. Hover over an expanded card to show the Latest user message button (a person and a speech bubble) between its documents and the answer: it opens the last message you wrote in that Agent Chat, in full and with its time, and Previous and Next step through your earlier messages. Only your own messages count; tool results, subagent prompts, notifications and context the client adds are left out. The text is read when you open it. Idle means the last turn finished, not that the Agent Chat is closed; a yellow Newly label marks the first minute after the turn ends. All projects shows the project name on each card where it is known. Only object IDs are sent to fetch names and details; Agent Chat content and names stay in VS Code.

Subagents count as the Agent Chat's work. For Claude Code, Codex and Copilot CLI, an Agent Chat whose main agent is idle while a subagent works shows running · 2 subagents, and the tab's running indicator stays on; a question waiting for you still shows as waiting for you · 1 subagent running. Main agent and the Subagents disclosure share one row. Their status labels are muted so the chat's overall Running stands out. Fold out Subagents for each agent's status, name or task and latest activity, working ones first. Documents and tasks a subagent changed appear on the Agent Chat's card, and answer text has its own shaded frame. A subagent that has been quiet for over 10 minutes without finishing is stale; it does not keep the Agent Chat running, and once the main agent has continued it no longer decides the Agent Chat's status. Copilot Agent Chats in VS Code's own chat storage show no subagents yet.

Your own markers help you sort the list. ✓ marks an Agent Chat done for yourself: it leaves Not Done but changes no task or document, does not stop the agent and archives nothing in Claude Code, Codex or Copilot. It is offered when the Agent Chat is not running or waiting for you. ☆ stars an Agent Chat. Hover over or focus the card to reveal ✎ at the end of its heading; it adds or edits a short note, shown directly after the title in pale yellow in a dark theme. Choose All, Not Done, Done, Starred or Reply: Not Done shows unmarked Agent Chats, every running or waiting Agent Chat, and Agent Chats marked done that have New activity since; Mark done again or Restore settles those. Agents starts at All each time VS Code starts, and keeps your choice until VS Code closes. A short message over the list confirms each marker change, with Undo right after ✓; a card that leaves the list you chose fades out. Opening, renaming or changing an Agent Chat's settings is not new activity. Markers are stored only in this VS Code profile (not synced, never sent to SpecViber), apply in every window and in All projects, and are kept when an Agent Chat's transcript is no longer found.

Reply attention helps you find chats that need an answer. Choose the Reply filter to show chats waiting for an answer or approval and visible Reply? hints, even if you marked the chat done. Combine it with text search, or search for reply or waiting for answer. Dismissed hints leave the filter immediately. Waiting for reply identifies an unanswered Claude Code question, and Approval needed identifies a Claude Code plan approval or a Copilot CLI permission request, a subagent's included. Reply? is a possible question or request for a reply in the last completed answer from Claude Code, Codex or Copilot CLI. It can be wrong, including for rhetorical questions: use × beside it to dismiss just that hint. A new answer can show a new hint. Text hints cover Swedish and English answers up to 8 KiB; longer answers and the older Copilot chatSessions format are left unclassified. Codex has no confirmed question signal in this version. Hints also appear in Compact and with answer previews hidden; they change no running status, personal done marker or card order. The Agents tab adds a yellow dot and separate confirmed/possible counts in its tooltip, alongside the green running indicator. All classification stays on this machine; dismissals store only chat/evidence identities and time in this VS Code profile.

Open followed by the agent icon reopens that card's Agent Chat. A single destination opens directly; otherwise a popup offers Codex · VS Code and Codex · Desktop app, or the corresponding client and destinations. Arrow keys select a destination; Escape closes the popup. Related earlier chats are available through History. The document/task panels use the same destination names. If the Agent Chat belongs to another project, choose Open that window or the app alternative. A successful handoff is quiet; a rejected IDE handoff offers the same Agent Chat in the app, plus the pickup prompt when the Agent Chat has an associated object.

Verify ▾ appears to the left of Open while you hover over or focus the card, and lets another agent check what this Agent Chat did; Open keeps its place. Copy verification prompt puts a review prompt on the clipboard: the Agent Chat's client, name, period and the documents and tasks it changed, your latest message in it, and instructions to check the work objects, the Git commits and the tests without changing anything. Paste it into a new Agent Chat yourself, preferably in another client. The agent's own final reply is left out on purpose, so the reviewer forms its own view first. Once the reviewer has reported, Copy latest agent reply gives you that reply to paste as a next message, so the reviewer can compare the agent's claims with its findings. The button confirms each copy. Your message and the reply are included only with Show Preview on. Nothing is sent anywhere.

Copilot · VS Code reopens Copilot Agent Chats in the VS Code window for their project, including Agent Chats stored by the Copilot CLI harness. This works from Agents and the document/task panels without the experimental setting.

A green dot on the Agents tab means at least one Agent Chat is running in the selected project scope. The tooltip gives the count. Search and object-history filters do not affect it; waiting and idle Agent Chats do not light it. Status follows changes while the navigator is visible, including on Docs / ASpecs, Tasks and All. Object names and answer previews are requested only when the timeline is visible. Monitoring stops when both the navigator and timeline panel are hidden.

A document or task the Agent Chat wrote to glows orange for up to 25 seconds after its latest write, on the expanded name or the folded ID badge. Each write restarts its own window; the glow stops when the turn finishes. Unfinished permission waits keep the remaining glow. With reduced motion enabled, the colour stays steady.

The agent menu on every task and registered document — navigator rows in Docs / ASpecs, Tasks and All, the document viewer's toolbar and the task and document panels — opens beside its button, with the same grouped layout as the web app: back to your own Agent Chat on this machine that changed the object, a new Agent Chat with the pickup prompt, or the object's Agent History (the timeline filtered to it) and Open in browser (everyone's Agent Chats on the web app's Agents page). The task header labels the button Open followed by the agent icon; other surfaces use Agents or the agent icon. This menu covers chat opening and history in both object panels. Agent History has ← Back to return to all chats. An ordinary document gets a document prompt, not the action-spec workflow. Opening the menu changes no status. Navigator rows also show ● Running or Waiting when one of your Agent Chats on this machine is working on the object or waiting for you; this comes from the timeline's own reading and is never sent anywhere.

Send to agent… in the viewer's text-selection menu and in a checklist item's ⋯ menu takes the selected text, or the item with its description, to a new Agent Chat or one of your earlier Agent Chats for the same document. The selection menu always offers it for readable text in a registered document, next to Comment, Add task and Highlight where those apply. The same popup shows the editable excerpt and optional instruction together, above the client choices. Your draft remains available if you cancel and reopen it. Where a client's link is verified to fill in text, the Agent Chat opens with the text in its input; otherwise the action says Copy text and open …, puts the whole text on the clipboard and opens the Agent Chat. An earlier Claude Code Agent Chat gets the text filled in when it is closed, and the text is also copied: Claude Code does not fill a chat that is already open, and says so. Nothing is ever sent for you: you review the text and send it yourself. The text stays on this machine, is never cut short, and works in read-only documents too; it does not change the item's state or who it is assigned to.