Changelog

Copy Markdown View Source

All notable changes to this project are documented here. The format is based on Keep a Changelog and this project adheres to Semantic Versioning.

[0.3.0] — Coordination, robustness & permissions

Multi-agent orchestration, long-session safety nets, and per-tool admission control — all opt-in, all offline-tested. 194 tests (was 162).

Coordination (pydanticAI levels 2 & 3)

  • ExAgent.Coordination.delegation_tool/2 — build a tool that runs another agent as a sub-task (agent-as-tool). The delegate's token usage is merged into the parent run's usage, so limits/cost cover the whole delegation tree. Accepts a static ExAgent.t() or a (ctx, args -> ExAgent.t()) builder.
  • ExAgent.Coordination.handoff/2 — transfer control between participants in a Session directly (BEAM-native programmatic hand-off), bypassing the turn policy.
  • ExAgent.Session.TurnPolicy.SupervisorPolicy — a coordinator participant alternates with each worker (supervisor, w0, supervisor, w1, …); registered as :supervisor / {:supervisor, supervisor: ..., workers: [...]}.

Robustness / cost

  • ExAgent.Compaction (behaviour) + Summary impl + Capability hook — shrink a long history to a summary + recent window on before_model_request, so a session stays within a model's context window.
  • ExAgent.UsageLimits — new tool_calls_limit (a batch that would exceed it runs nothing) and max_budget_cents (halts via an :estimate_cost function). New check_tool_calls/3; cost checked in check_before_request.
  • ExAgent.CostGuard.estimator/1 — turn a pricing map into the (usage -> cents) function for max_budget_cents. No pricing table baked in.
  • Anthropic prompt caching: cache: true on ExAgent.Models.Anthropic adds cache_control breakpoints to the last system block and last tool.

Permissions

  • ExAgent.Permissions — per-tool :allow / :ask / :deny with glob rules (last-match wins, fail-closed :ask). Wired into run/3 via :permissions and :approve; a denied tool returns a "not permitted" message to the model.

Core (general, backwards compatible)

  • Tools may now contribute token usage by returning {:ok, value, %ExAgent.Message.Usage{}}; the loop merges it.
  • ExAgent.PubSub.Phoenix validated end-to-end against a real LiveView PubSub.

[0.2.0] — Stateful runtime, sessions & persistence

The first release with the full layered runtime on top of the existing one-shot core. Nothing one-shot broke — the new layers are opt-in. See DESIGN.md for the architecture and ROADMAP.md for the phased plan.

Layer 1 — Stateful agent (ExAgent.Server)

  • ExAgent.Server — a supervised, long-lived agent that preserves conversation history, accumulates usage, threads stateful models across runs, and emits events. API: chat/3, send_message/3 (async → event), stream/3, steer/2, abort/1, set_model/2, history/1, usage/1, health/1.
  • ExAgent.AgentSupervisor — DynamicSupervisor to start/stop agents by id.
  • Runs execute under ExAgent.TaskSupervisor, so a Server keeps answering abort/1, health/1 and backpressure (:busy / :queue_full) during long runs.
  • reset/1 — clear the conversation history and accumulated usage (e.g. a chat UI's "clear" button); only when idle.

Layer 2 — Persistence (ExAgent.Store)

  • ExAgent.Store behaviour + ExAgent.Store.ETS (dev/test default).
  • ExAgent.Server.Snapshot — a strictly JSON-serializable checkpoint (history + usage + metadata). By construction it never carries pids, secrets or tool closures; the live model/tools come from an app-supplied template on restart.
  • ExAgent.Server checkpoints after every run and rehydrates history/usage on restart (store: :ets). Survives supervised-agent crashes.

Layer 3 — Session & coordination (ExAgent.Session)

Cross-cutting

  • ExAgent.Event — versioned, serializable event envelope (the UI/runtime contract; distinct from :telemetry).
  • ExAgent.PubSub behaviour + None / Local (Registry) / Phoenix (dynamic, no hard dependency) impls. Agents publish on "exagent:agent:<id>", sessions on "exagent:session:<id>".

Core additions (backward compatible)

  • ExAgent.run/3 accepts :on_event (loop event sink, no-op by default), :prepend_instructions, and :run_id. The result map now also exposes :model (the possibly-updated model, so stateful models thread across runs).

Docs & examples

  • DESIGN.md, ROADMAP.md. New examples: examples/stateful_agent.exs, examples/multi_agent_session.exs. ex_doc module groups by layer.
  • 159 tests (was 96), stable across repeated runs.

[0.1.1]

  • Fix structured-output JSON Schema generation for embeds_many / embeds_one.

[0.1.0]

  • Initial release: agent loop, providers (OpenAI / Anthropic / Z.AI / OpenRouter / Test), deftool-derived tool schemas, Ecto structured output with retry, streaming, capabilities, UsageLimits, telemetry, message-history serialization.