High-level flow
- An agent starts a background command.
- The runtime launches and tracks that process with ownership metadata, timeout rules, and bounded log capture.
- Humans can watch process activity in the app and intervene when needed.
- Agents can query process state during execution.
- When the process exits, the runtime injects a structured result back into the owning thread.
- If that owner is idle, the runtime resumes it so the result is not missed.
Observability for humans and agents
- Humans get process visibility in the desktop Process Monitor: command, pid, owner, status, start/end time, and log tails.
- Agents can inspect live background work from tools, including running process lists and status snapshots.
- In team contexts, members can inspect currently running team processes started by other members, so work is visible across the team instead of isolated to one thread.
Control and safety
- Humans can stop processes directly from the app when a command is stuck, unsafe, or no longer useful.
- Agents can request process termination through process tools.
- Runtime limits (timeouts, log caps, ownership checks, and permission policy) keep background execution predictable and fail-closed.
Auto-injection and wake-up behavior
When a background command completes, its final state and output are injected into the owner thread as a synthetic tool result. This means:- The thread keeps full context without manual copy/paste.
- Follow-up reasoning can use the real command output immediately.
- Idle agents are resumed so the completion signal is acted on.
Why this matters
The Process Manager gives long-running automation the reliability of a job system with the usability of chat:- commands keep running even when a turn should continue,
- everyone can observe what is running,
- and results reliably flow back into the right place.