Read-only authoritative activity for an existing live agent instance

[TaoSama]

For safe maintenance of an integration using existing terminal sessions, is there a supported read-only way to obtain authoritative activity for the exact live agent instance, rather than its screen-derived status?

With Herdr 0.9.5/protocol 22, the advertised agent/session snapshot and foreground-process APIs do not declare complete background or child coverage. Related reports #5004, #5261 and #5889 show why display idle and a historical session ID cannot establish that coverage. I am not reporting these as a new reproduced bug or proposing a detector rule.

The needed distinction is complete versus unknown for an existing session: native runtime identity/generation, all children and background tasks, pending interactions/compaction, and a bounded fresh revision. Reading another app-server or starting/resuming a replacement session would not cover the original instance. Installing hooks after a session starts must not imply complete lifetime coverage.

Is an existing maintainer-owned interface or accepted issue intended to provide this, including for Claude and embedded Codex sessions? If the native runtime lacks such a contract, confirmation that these sessions remain unsupported would also help consumers preserve unknown correctly.

I am following this dependency for the downstream integration and will keep its collection fail-closed. I am not claiming an upstream implementation owner, submitting an unsolicited implementation PR, or asking for contributor access.

[Gianni300x]

Hi! I checked the Herdr documentation, particularly the Agents and Socket API sections.

From what I can tell, Herdr exposes agent state, session snapshots, and event subscriptions, but I couldn't find a documented guarantee that these provide authoritative, complete activity coverage for an existing agent instance, including background tasks and child processes.

The distinction you're making between an observed idle state and confirmed inactivity seems important.

For a downstream integration, I agree that treating incomplete or stale observations as unknown is safer than assuming an agent has finished.

The remaining question seems to be whether there is a maintainer-supported interface that provides those completeness guarantees for Claude Code and embedded Codex sessions.

I couldn't verify such a contract in the public documentation, so maintainer confirmation would be helpful.