[idea] Structured task, session, model, and policy evidence for agent automation

[sunilgupta01]

I use Herdr as the runtime for a launcher that opens the same logical review
crew in multiple workspaces. Herdr 0.8.2 gives the client stable workspace,
tab, pane, process, cwd, readiness, and sometimes provider-session identity.
Those facts are useful, but they are not enough to decide whether an existing
interactive agent is safe for a new task.

The missing structured evidence is:

• a client-supplied task/run identity or ownership token bound to an agent;
• a supported fresh-session/reset operation that returns the new session ID;
• effective provider/model identity when an integration can observe it, with an
  explicit unknown state otherwise;
• capability/evidence fields for read-only and child-process restrictions; and
• a compare-and-close operation or revision/token check for owned cleanup.

Today agent names are globally unique while workspace identity is separate, and
agent_session can be absent. A safe automation client therefore has to derive
its own workspace/run-scoped names, keep an external receipt, refuse legacy or
unattested reuse, and quarantine any lane whose required model or execution
policy is unobservable. That works, but it prevents Herdr from being the
authority for freshness and policy compliance.

Would a small versioned automation contract fit Herdr's direction? The minimum
useful shape for this workflow would be opaque client run metadata on start,
structured session/model/policy evidence on get/list, and conditional cleanup
against the returned identity/revision. The metadata would not need to affect
the TUI or interpret any repository-specific task format.