Steer the agent while it works

A message you send while an agent is working now goes into the running turn instead of waiting for it to end. Its bubble says “Steering” until the agent picks it up. The message then appears in the conversation at the point the agent read it, and the rest of the reply answers it.

Prokop agents read it after their current tool step, Claude Code between tool calls, and Codex through its own steering request. If the turn ends first, or the agent can’t take it (a Claude Code or Codex Goal run, or a stopped turn), the message stays in the queue and is sent after the turn, so it is never lost.

Prefer the old behavior? Set Settings → Appearance → Messages sent while the agent works to Queue. The choice is kept per device. A queued message has a Steer button, and pressing Enter in an empty message box steers the oldest one.

Switch harness mid-conversation

When Claude Code or Codex runs out of usage, pick a model from another harness in the session’s model picker and keep working in the same session. Before, the first message locked the harness.

The new harness starts a fresh thread with a summary of the conversation: the first request, the last six turns with one line per tool call, and the session ID so it can look up earlier details with session search. Nothing from the earlier turns runs again. The conversation shows “Switched to Codex CLI · gpt-5” where the switch happened.

Switching works while the session is idle, with no queued messages, pending approvals, running Goal, or Compact. After a switch, Fork is unavailable, and Edit and Undo are unavailable in Claude Code and Codex.

Pages, diagrams, and math in replies

Agents can show pages. Prokop, Claude Code, and Codex agents can show a chart, table, diagram, or mockup as a page in their reply. It follows your light or dark theme, grows to fit its content, and opens in a new tab from its corner button. Pages run sandboxed, with no access to your login or stored data, and can be up to 200 KB. This comes from a new built-in MCP entry, Prokop Visuals, which is on without setup. Turn it off, or turn off its render_html tool, in MCP settings for every workspace or for one.
Mermaid diagrams. A ```mermaid block is drawn as a diagram once its closing fence arrives. A corner button switches between the diagram and its source.
Math. Math between $$ signs and ```math blocks render with KaTeX. A single $ stays text, so “$5 and $10” is not turned into math.
Codex images show inline. Images Codex generates show under their row, also in older turns. Click one to open it full size. Prokop keeps its own copy with the session.

Terminal

Reorder tabs. Drag a terminal tab to a new position. The order is kept per workspace on this device.
Terminal on phones. The terminal sheet stops above the on-screen keyboard, so typing never hides the prompt. A key row adds ^C, Esc, Tab, and arrow keys without opening or closing the keyboard.

Fixed

Codex image generation no longer breaks the session. An image larger than about 1.5 MB dropped the connection to Codex mid-turn and left the session stuck waiting to reconcile.
File links in replies open the file. Clicking a file link in a reply opens the file preview instead of a broken tab. Files the preview can’t show, such as spreadsheets and PDFs, have a Download button.
Stop responds at once. The session switches to idle immediately instead of waiting about 2 seconds for Claude Code to shut down. A message sent in that moment runs after the stopped turn ends.
Following the conversation stays at the bottom, even after sending a long message or receiving a large tool output.
Claude Code and Codex read their own skills without asking. Reading skills in their own folders, such as ~/.claude/skills or ~/.codex/skills, no longer asks for outside-workspace access. Writing there still asks, and so does reading anything else in those folders.
Claude Code web results stay folded to one line, with the full results under Show raw data.
Opened attachments can’t run scripts. An attached HTML page opened from the conversation could run as Prokop and read your API token. Attachments now open sandboxed.

Update

Run this on the machine hosting your Prokop server:

prokop update