Changelog

All notable changes to req_managed_agents_host are documented here.

v0.2.0

Transcript persistence: providers that emit SessionResult.transcript (req_managed_agents >= 0.10, e.g. Providers.Local) get it stored under a sibling {:transcript, external_id} Store key and re-injected as history: on reattach — local conversations now survive process crashes, idle-detach, and full BEAM restarts.
Server-held providers (Claude Managed Agents, AgentCore) are unaffected: no transcript emitted, nothing stored, opts unchanged.
Requires {:req_managed_agents, "~> 0.10"}.

Initial release: a durable, single-node session host over req_managed_agents.

ReqManagedAgents.Host.send_message/3 — the whole public surface. Finds, starts, or reattaches the live session for a caller-supplied external id, runs one turn to a terminal via ReqManagedAgents.Session.run/2, and returns the RMA SessionResult.
ReqManagedAgents.Host.Config — validated, struct-first host configuration (:provider, :handler, :store required; optional :agent/:environment handles, :idle_timeout_ms, :timeout_ms, :metadata).
ReqManagedAgents.Host.Store behaviour, with two implementations:
- Store.ETS — process-local, in-memory; gone when its owner process stops.
- Store.DETS — file-backed; the locator survives a full BEAM restart.
ReqManagedAgents.Host.Locator (+ Locator.Record) — the durable external_id -> upstream identity map that backs reattach and idle-detach.
ReqManagedAgents.Host.SessionServer — a per-external-id GenServer driving one turn at a time. Idle-detaches after idle_timeout_ms; its child spec is restart: :temporary, so both idle-detach and a crash simply end the process — the next send_message/3 re-creates it and reattaches through the Locator rather than being auto-restarted in place.
ReqManagedAgents.Host.SessionSupervisor — a DynamicSupervisor holding one live session per external id; killing one server never affects a sibling.
An OTP application (the mod: entry in mix.exs) starts the session Registry and the SessionSupervisor.