v0.1.20

Added

• Message-tree conversation history. History now lives in a tree the engine
  owns and sends on each worker decision, not in worker state. The worker returns
  the transcript as it should look and the engine reconciles it, forking a branch
  where it diverges. This enables edits, regenerations, plan/execute, and prompt
  compaction.
• The worker decision includes a count of effects still running, so a worker knows
  when a step is done without scanning the list.
• A language-neutral protocol reference (docs/07-protocol.md), so workers can be
  written in any language.
• New assistant-ui-cloudflare-starter example: an assistant-ui chat on a
  Cloudflare Worker, streaming from AG-UI.

Changed

• Breaking (SDK): the agent API is rebuilt. An agent is now a single decision
  function (the built-in tool/sub-agent loop, or your own returning plain action
  objects), replacing the middleware/builder API. Models are configured inline for
  the server to run, or supplied as an adapter your worker runs.
• Breaking (wire): one unified effect protocol. Tool, model, and sub-agent
  calls share a single request/settle vocabulary instead of a message type per
  kind, and sub-agent results are reported as themselves, not as tool results. The
  settle endpoint and its SDK/runtime methods were renamed to match.
• Breaking (wire/SDK): parallel model calls. A worker can issue several model
  calls from one decision that settle independently; the one-at-a-time restriction
  is gone. The engine still paces server-run calls; run them on the worker for real
  overlap.
• Breaking (SDK): tools. A tool's execute receives the decision it runs under,
  not a separate context; async tools are flagged with an option, not a sentinel;
  tool state lives in your own store; and client tools may omit execute.
• Breaking (SDK): streaming. A worker-run model streams tokens through one
  delta shape; the intermediate stream-part type and its wrapper are gone.
• Breaking (wire): the decision request's end-user field is renamed from
  owner to identity, matching the SDK and docs.
• Workers no longer send a span with submissions; the engine owns tracing.
• Model-call actions default their retry and streaming options, and must state who
  runs the call.
• The AG-UI /run endpoint forwards the client's full transcript, which the engine
  reconciles into the tree.

Removed

• Breaking (SDK): the middleware system and the Substructure.agent factory;
  the AI and OpenAI adapter classes become function equivalents.
• The substructure new command and the templates/ directory (starters moved
  into examples/).
• The examples/ pnpm workspace; each example is now a standalone npm project.

v0.1.19

Added

• substructure new [template] [dir] scaffolds a project by copying a starter
  template from the monorepo. With no template it shows a picker (or lists
  templates when non-interactive).

v0.1.18

Removed

• The CLI no longer prints zero-balance warnings when targeting an app. App
  balances are still shown by apps list and apps show.

Changed

• Interrupts and cancels now fully fence the session: pending work is voided,
  late results and client actions can't wake a paused session, and decision
  requests raised mid-pause queue until resume.

v0.1.17

Added

• openaiAgent factory in the OpenAI adapter: converts an @openai/agents
Agent (or OpenAIAgentSettings) without a second new, e.g.
  openaiAgent(new Agent({ ... })).
• Interrupts record the issuing caller's origin; resuming requires equal or
  higher privilege.
• Clients can interrupt and resume their own sessions via the client API.
• New interrupt worker action lets an agent pause awaiting external input.
• AG-UI interrupt-aware run lifecycle: interrupt outcomes and resume[].

v0.1.16

Added

• Native AG-UI protocol support: endpoints that stream an agent turn as AG-UI SSE
  events, with live token and reasoning streaming.
• Reasoning controls on LlmRequest (reasoning: effort or max_tokens,
  plus exclude and enabled), passed through to providers that support them.
• llmToolLoop takes a generator: a worker-side provider generator that runs
  the LLM call on your worker, or serverGenerate to let the Substructure
  server's configured provider make the call.
• Anthropic adapter (@substructure.ai/sdk/adapters/anthropic): call the
  Anthropic Messages API from a worker via the anthropicGenerate generator.
• AI SDK adapter (@substructure.ai/sdk/adapters/ai): run an existing Vercel AI
  SDK agent on Substructure via ToolLoopAgent, or drive the loop directly with
  the aiGenerate generator.
• OpenAI adapter (@substructure.ai/sdk/adapters/openai): run an @openai/agents
Agent on Substructure via OpenAIAgent, or drive the loop directly with the
  openaiGenerate generator.
• The server and embedded runtime can start without an LLM provider configured.
• $SUBS_API_TOKEN sets the CLI's bearer token without a login — for targeting a
  self-hosted server with auth enabled (a serve --dev server needs no token).
• substructure login now stores tokens per server URL, so you can stay logged
  in to several servers at once (e.g. cloud and a staging deploy); each command
  sends only the token for the server it targets.

Changed

• Breaking (CLI): subcommands are now top-level — substructure cloud <cmd>
  becomes substructure <cmd> (e.g. substructure sessions list), and
  substructure local start becomes substructure serve.
• Breaking (CLI): the credentials file is now keyed by server URL with no
  migration of the old single top-level token; existing logins must re-run
  substructure login.
• Breaking: session events identify a session's end user as owner (was
  identity); consumers reading raw events should read owner. No alias for the
  old key.
• Breaking: Caller::System is now System { tenant_id } (was a unit
  variant); crate consumers constructing or matching it must supply a tenant.
• Breaking: the llmLoop middleware (and agent.llmLoop) is renamed
  llmToolLoop to make the llm-and-tool loop it drives explicit.
• Breaking (worker protocol): a turn's tool and sub-agent results are
  delivered as a single batched effects.complete trigger, replacing the
  per-effect tool.result, sub_agent.turn.complete, and sub_agent.error
  triggers; spawn.sub_agent now carries tool_call_id. Handled transparently by
  SDK workers.

Fixed

• OpenRouter responses no longer drop image outputs from the stream.
• The next LLM call in a turn waits for all of the turn's effects to finish, so
  turns with multiple sub-agents or a mix of tools and sub-agents no longer call
  the model with partial results. Sub-agents run concurrently.

v0.1.15

Added

• Changelog with automated release notes.

Changed

• Agent state is serialized as JSON automatically; chains no longer need an
  explicit .use(agent.jsonState()). Existing chains that still include it
  keep working, and the wire format is unchanged.

Fixed

• Tool-result messages (message.new with role: "tool") now carry the tool
  name instead of null, so consumers reading only the message stream (e.g.
  reconstructing a transcript from session events) no longer have to correlate
  tool_call_id back to the originating tool call to recover the name.