Chat Monitoring
Chat monitoring in ADHDev means you can follow active agent sessions from the dashboard instead of watching each tool separately.
What ADHDev Monitors
ADHDev can surface conversation activity from several different kinds of agent runtimes:
- Native IDE chat from supported desktop IDEs
- Extension chat from agent extensions running inside a host IDE
- CLI sessions captured through the hosted PTY runtime
- ACP sessions from providers that expose structured protocol events
These all show up as sessions in the dashboard, even though the underlying transport is different.
What You Can Do
Monitoring is not read-only. From the dashboard you can usually:
- follow live output while the agent is generating
- load older history for the active session
- send the next message without returning to the original tool
- queue a message while the agent is still generating, and send it immediately with Send now
- review approval-required turns and respond from the dashboard
- switch a CLI session between terminal and chat rendering
- open Remote View for supported IDE sessions
The exact controls depend on the provider. Some surfaces expose model or mode controls, while others are mostly stream-and-reply.
Tool Calls And The Activity Toggle
Tool calls (file edits, shell commands, and similar actions the agent takes) render as ordinary chat bubbles with the actual tool name — for example "Write", "Bash", or "run_command" — rather than a generic "Tool" label.
The dashboard's activity toggle controls the noisier rows: terminal output and thinking/reasoning text. Tool-call bubbles are shown by default regardless of the toggle; turning activity off hides only terminal/thought rows, so you can quiet the stream without losing the record of what the agent actually did.
Queued Messages
A message you send while the agent is still generating is queued rather than dropped:
- it persists across a page reload
- more than one message can be queued at once
- queued messages stay pinned to the bottom of the conversation
- a queued message can be cancelled before it sends
- Send now hands the message straight to the agent's own mid-turn input queue where the provider supports it, instead of interrupting the current turn
Live Sync And Replica Status
Chat history is delivered through a durable transcript replica in addition to the live tail. If the replica falls behind or a connection briefly drops, the dashboard can show a "Live sync is having trouble" banner — but only after a short grace period, so a quick reconnect (a tab resume, a brief network blip) does not flash the warning unnecessarily.
How Sessions Appear
The dashboard uses the same monitoring model across IDE, CLI, and ACP sessions:
- active sessions stay in the main tab strip
- lower-priority sessions can be moved into Hidden tabs
- sessions that need attention or have unread completed work appear in the Activity inbox
- older or resumable sessions can be reached through History
This is why ADHDev monitoring feels closer to a workspace than a simple notification feed.
If a machine has exhausted its auto-reconnect budget, you can reconnect it directly from the chat header (next to the bell/session-info icons) without navigating to the Machines page.
CLI Monitoring
CLI monitoring is different from IDE monitoring because the terminal itself is part of the product surface.
For CLI sessions, ADHDev keeps the hosted PTY runtime and exposes:
- full terminal output
- reconnect-friendly scrollback
- approval prompts and tool output
- a terminal/chat toggle for the same runtime
If a CLI session survives but the dashboard loses track of it, recovery usually happens through the machine's Hosted Runtimes tools rather than by creating a brand new chat.
Approval And Attention States
ADHDev treats waiting states and completed states as first-class signals.
Examples include:
- approval requests
- long-running generations
- finished tasks you have not reviewed yet
- disconnected or interrupted sessions that need manual recovery
Those states are what drive the activity inbox and notification surfaces.
Limits And Caveats
- Not every provider exposes the same amount of metadata or control
- IDE monitoring depends on the control surface the IDE makes available
- CLI and ACP providers can exist in the built-in inventory before they are fully validated
- Some sessions are better described as monitored and recoverable rather than fully controllable
For practical expectations, check Compatibility & Caveats and Supported Providers.