Conversation view

Every claude session can be shown two ways: as its terminal (the default) or as a conversation — messages, the agent’s thinking, and tool results laid out like a chat, with a box at the bottom to send the next message.

Shell sessions have only the terminal view.

Switching views

Use Show conversation view and Show terminal view on the session. The choice applies to all agent sessions at once and is remembered across restarts, so new sessions open in the view you used last.

Switching never restarts anything. Both views show the same running claude process; the terminal keeps running underneath while the conversation is on screen.

Where the content comes from

The conversation view is built from the record Claude Code itself writes for each conversation, in ~/.claude/projects, read as it grows. Whether the agent is working or waiting comes from events Claude Code reports to spekterm while it runs.

spekterm never reads the terminal screen to build this view. Screen output is made for people, not programs; reading it breaks whenever the CLI changes how it draws. The terminal view stays available for anything the conversation view cannot show.

When the view opens a long conversation, it loads the most recent part and says Earlier messages are not loaded. While it catches up you see Catching up…. If the record cannot be read at all, the view says Conversation is unavailable instead of showing an empty conversation.

Sending a message

Type in Message the agent and press Send. spekterm only lets you send while the agent is waiting for your input:

• Working… — the agent is busy. Wait for it to finish, or interrupt it from the terminal view.
• The agent is asking to use a tool, or is waiting for a choice. The view shows what it is waiting for. Answer it in the terminal view — the conversation view does not send free text into a permission prompt or a menu.
• The agent’s state is unknown. This happens right after a session starts, or when Claude Code is not reporting events. The view refuses to send rather than guess; switch to the terminal view to see what the agent is doing.

What you send goes to the agent as plain text. It is never interpreted as terminal control sequences.

A message you sent appears as Not confirmed yet until it shows up in Claude Code’s own record. If it never does, the view tells you to check the terminal view — the message may not have arrived.

Hibernated sessions

A dormant or hibernated session shows Session is hibernated. Waking resumes the conversation. Press Wake (or Enter) to start it again; the conversation continues where it stopped. See Terminals and sessions.

When to use which

• Use the conversation view to read what an agent did, follow a long answer, or send the next instruction without terminal noise.
• Use the terminal view for permission prompts, menus, slash-command pickers, and anything interactive — and whenever the conversation view tells you it cannot help.

What it needs from Claude Code

spekterm adds its own settings to each claude session it starts, so Claude Code reports when it starts working and when it waits. Your own Claude Code settings, hooks, and status line keep working; spekterm adds to them and does not change your settings files. Agent sessions started outside spekterm are not affected.