Changelog

Find a version

All versions All versions forward

Version 0.18.0

Every @aparte/* package ships at this version (they are released in lockstep).

Minor Changes

• 8f2d28e: New: run the agent loop on your server. createAparteAgentHandler runs it next to your tools and streams it; createAgentStreamRunner plugs AparteClient into it — tool rows, Stop, the approval panel and your approval policy work as they do in the browser, and a user can leave in the middle of an approval — another conversation, a closed tab — and answer it when they come back.

  Until now a server with its own tools could only drive the chat display-only: the message, the segments, closing the turn and four lifecycle events, by hand. The handler pauses the loop on a call that needs a person, or on a tool the browser runs (ask_user), and ends the stream with a signed continuation ticket; the page asks — the same panel, the same policy — and resumes with the answer, in the same bubble, under one pair of lifecycle events. Nothing is stored on the server: the paused call lives in the conversation you already save (AparteToolCallSegment.deferred) — saved the moment the run pauses, since the reply now dispatches aparte-message-paused and the conversation controller persists on it — and a conversation that loads with one resumes on its own — the conversation controller now dispatches aparte-conversation-loaded, and client.resumeDeferred() is there for a host that loads conversations another way. A message sent instead of an answer closes the paused call.

  The server decides which calls need a person; the page’s policy answers when it has an opinion, and otherwise the person is asked. The ticket makes the server run only the calls it emitted, for the reply it paused; it does not make a tool’s input trustworthy, and it is not single-use unless you pass onTicket. A server tool checks who is calling with the request it gets as context.request (AparteToolContext.request, new). A toolChoice naming a tool, or a history ending on unanswered calls without a ticket, is refused with a 400, and a server tool’s own error text stays in the server’s log. An approval payload is dropped unless the tool accepts one. Guide: Run the agent loop on your server. _@aparte/core

• 3776dfc: The model now remembers what its tools returned: each turn’s history carries the tool calls and their results, not only the text. If you serialize the history yourself — a backend of your own that does not go through an aparté provider or toModelMessages — it now contains assistant messages with toolCalls and role: 'tool' messages, the shapes a tool round trip already used inside one turn: map them, or drop both. Through an aparté provider there is nothing to do. Requests grow on tool-heavy conversations; @aparte/plugin-compaction now counts those calls when it measures a conversation, and summarises long ones as before.

  Until now AparteClient sent user and assistant text alone, so a read_file result was on screen and gone from the model’s context one turn later — “what did the file say?” had no answer. A reply is now sent in the order it was shown: the text before some calls and those calls as an assistant message, each call’s result as a tool message, the text after them as the next assistant message. Every call is answered, because providers reject a call left without one: a refused call sends its reason, a failed one its error line, a stopped one “The tool call was stopped before it ran.” Send, retry and edit build the history through the same function, so they cannot disagree. The compactor’s selector used to count a tool call as free (it has no content); it now costs what its arguments and result send the model. _{@aparte/core, @aparte/plugin-compaction}

• 65845fe: New: runStreamAgent can pause on a tool call it cannot finish and resume it later. Pass deferral and a call that needs a person’s decision, or a tool marked external, ends the run on run-suspended instead of waiting; a later run whose history ends on those calls, with their answers in deferral.outcomes, picks up where it stopped. If you switch exhaustively over StreamRunEvent (an adapter of your own), add a case for run-suspended; it only occurs when you pass deferral.

  Nothing changes for a caller that does not pass deferral: the approval resolver still answers in place. The run that pauses reads the model’s reply to its end, so every call of that reply is pending at once and a client can answer them in one round trip; nothing runs after the pause point, and a call the gate denies is refused, never deferred. A resumed run answers its pending calls through the same code a streamed call takes, and emits no tool-start for them, since their rows exist already. unansweredTail(messages) returns the calls a history leaves unanswered. This is the loop half of running the agent on a server; @aparte/core adds the server handler and the browser runner. _{@aparte/engine}

Patch Changes

• 81661e9: Stopping a reply no longer prints an uncaught AbortError in the console when the stream had already failed.

  An aborted fetch errors its response body, and the Stop handler then cancelled the reader. reader.cancel() returns a promise that rejects on an errored stream, and it was wrapped in a try that only catches synchronous errors, so the rejection went unhandled. It is caught now. _@aparte/core

_{Version-only bumps (no changes of their own): @aparte/provider-ai-sdk, @aparte/provider-openai-compat, @aparte/provider-scenario, @aparte/provider-transformers, @aparte/plugin-approval, @aparte/plugin-artifacts, @aparte/plugin-ask-user, @aparte/plugin-marked, @aparte/plugin-model-selector, @aparte/plugin-shiki, @aparte/plugin-streaming-markdown, @aparte/plugin-titler, @aparte/angular, @aparte/react, @aparte/svelte, @aparte/vue, @aparte/locale-fr, @aparte/docs-mcp.}