v0.11.0: 0.11: zastavení odpovědi, timeout podle aktivity, subagenti

- user.interrupt přes ConversationTurnMonitor / ChatEndpoint::cancel, přerušená odpověď se uloží (interrupted) a konverzace pokračuje ve stejné session
- místo maxPolls idleTimeoutSeconds (dotaz na stav session) a maxTurnSeconds, maxToolCalls po vláknech
- čtení událostí od nejnovější s stránkováním (dlouhé session nad 1000 událostí)
- multiagent coordinator {type: self} (AgentProvisioner subagents), karty subagentů ve widgetu
- widget: tlačítko Stop, živé subagenty, timeout podle aktivity workeru

v0.10.1

Widget: po příchodu odpovědi odskočí na její začátek, ne na konec

v0.10.0

Model a přemýšlení jen u zpráv: konverzace běží na volbě své poslední…

v0.9.2

Jméno odesílatele u každé zprávy, kterou nenapsal zakladatel konverza…

v0.9.1

Kontextové okno ve výběru modelu, ukazatel kontextu podle zvoleného m…

v0.9.0

Shared conversations, model switching per message, robust render tool…

v0.8.0: Model selection, thinking effort and conversation rating (0.8)

ModelSelection lists the models and effort levels a user may pick for a
conversation. The choice is stored on the conversation and sent as a
per-session override of the agent (agent_with_overrides); it is locked after
the first message, because a session cannot switch models. The context gauge
uses the window of the chosen model.

The owner can rate a conversation (thumbs up / down + an optional comment) via
ChatService::rateConversation() and ChatEndpoint::rate(); a read-only view
shows the rating.

The widget gets the model and effort selects, the rating box and `params`
(+ setParams()) for extra request parameters.

BC break: ManagedAgentsClient::createSession() has a 4th parameter $model,
ConversationInterface new accessors and the conversation table five new
columns - see "Upgrading to 0.8".

v0.7.0: Provision the agent lazily before a new session

AgentProvisioner::resolver() creates the agent when the schema has none and updates it when the system prompt or the tools changed, detected by a fingerprint stored on the mapping (AgentTrait::$fingerprint, new nullable column). An unchanged agent costs no API call, so a deploy needs neither an API key nor a call to Anthropic, and a broken key of one tenant fails only that tenant's chat. ensureProvisioned() keeps updating unconditionally unless force: false.

v0.6.1

Allow adt/doctrine-anonymization 0.3

v0.6.0: Client: API key from a provider asked on every request

GuzzleManagedAgentsClient takes either a fixed key or a Client\ApiKeyProvider, for
applications where every tenant enters its own key. The x-api-key header is now set
per request instead of in the HTTP client defaults, so a long-running queue consumer
picks up a changed key without a restart.