Changelog
Release notes for aparté 1.8.0: what changed in each @aparte package, and what to update.
On this page
Version 1.8.0
Every @aparte/* package ships at this version (they are released in lockstep).
Minor Changes
-
b85cd5e: Tool calls run one after another again unless the tool says
parallel: true; a call waiting for an approval holds the calls after it;createAparteAgentHandlerandAparteClienttakeparallelTools.ts// A coding agent: reads run together, the edit and the command keep their order. aparteGlobalConfig.registerTool( { name: "read_file", description, inputSchema, parallel: true }, readFile, ); aparteGlobalConfig.registerTool( { name: "edit_file", description, inputSchema, needsApproval: true }, editFile, ); aparteGlobalConfig.registerTool( { name: "run_command", description, inputSchema }, runCommand, ); // A tool from an MCP server: it may run beside others when it says it only reads. const parallel = mcpTool.annotations?.readOnlyHint === true; // Every call one after another, `parallel` tools too: new AparteClient({ parallelTools: false }); createAparteAgentHandler({ ...options, parallelTools: false });Do you have to touch your code?
- If you relied on 1.7.0 running the calls of one reply at the same time: give those tools
parallel: true(onAparteTool, on a server tool’stool, or onStreamToolConfigforrunStreamAgent). - If you time a call from its
tool-start:tool-startnow comes when the model writes the call, before it runs, so a call that waits its turn counts its time in the queue. A queued call (queued: true) runs from itstool-running. - If you
switchexhaustively onStreamRunEvent['type'](orAparteStreamRunEvent['type']): the newtool-runningevent needs a case.
Fixes #202.
Since 1.7.0 every call of one model step ran at the same time. For a coding agent that broke what the model meant:
edit_filethenrun_commandraced, and with an approval policy the command (allowed) ran while the person was still being asked about the edit, on the file as it was. A server loop could not opt out at all.- A call keeps its place by default. A call of a tool that says nothing starts once every call made before it in the same reply has settled, and the calls made after it start once it has: the order every tool had until 1.7.0, for anyone who does nothing. Calls of tools that say
parallel: truerun together, between the calls that keep their place. This is the rule coding agents follow — read-only tools together, writes and commands one at a time — and it maps onto MCP’s annotation:parallel: tool.annotations?.readOnlyHint === true. - A call waiting for an approval holds the calls after it,
parallelones included. While a call waits for a person’s answer (or the policy’s), the calls the model made after it in the step do not start; the calls made before it keep running. An approval lets them start; a refusal skips them all, each answered “Not run: an earlier call of the same step was refused.” (by: 'earlier-refusal'). This changes 1.7.0, where the calls that needed no approval ran meanwhile. - A call waiting its turn shows where the model wrote it, queued. Every call gets its
tool-startas it is read, so its row stands before the text the model wrote after it. One that cannot run yet carriesqueued: true, and the new run eventtool-running { toolCallId }says when it goes on (never sent for a call that went on at once). A queued call a refusal reaches is rejected as an earlier refusal; one a Stop reaches ends astool-aborted('stopped'). The server loop’s NDJSON carries both, andcreateAgentStreamRunnerhands them on. AparteToolCallSegment.queued(true, besidestatus: 'pending'— thestatusunion is unchanged): the adapter sets it on a queuedtool-startand removes it once the call goes on or ends. The row reads Queued (new locale keytoolQueued, in all nine@aparte/localesbundles) with no spinner and no shimmer, its name in the muted ink, and the row carriesdata-queued. Live only: a restored conversation never keeps it. The turn summary names the step that runs, not one that waits, and so does a sub-agent’s line.parallelToolskeeps its meaning — run together the calls whose tools allow it,trueby default — andfalsestill runs every call one after another.createAparteAgentHandlerandAparteClientnow take it and pass it to the run, as they passmaxTurns.parallelis new onAparteTooland onStreamToolConfig;AparteClientandcreateAparteAgentHandlerhand a registered tool’sparallelto the loop, as they hand itsneedsApproval. A sub-agent’s run follows the same rule, through its run’s tool configs.@aparte/provider-scenario’sscenarioSubagentToolsaysparallel: true, so two delegations of one reply still run side by side.
What stays as in 1.7.0: the history is in call order, a person is asked about one call at a time, a Stop or the turn limit stops every call of the step, and with
deferrala call that pauses does not stop the calls already running (the calls made after it are queued with the pause, and the resume runs them in order).@aparte/core,@aparte/engine,@aparte/provider-scenario,@aparte/plugin-subagents,@aparte/locales - If you relied on 1.7.0 running the calls of one reply at the same time: give those tools
Version-only bumps (no changes of their own): @aparte/runner-ag-ui, @aparte/runner-n8n, @aparte/provider-ai-sdk, @aparte/provider-openai-compat, @aparte/provider-transformers, @aparte/plugin-approval, @aparte/plugin-artifacts, @aparte/plugin-ask-user, @aparte/plugin-cards, @aparte/plugin-chartjs, @aparte/plugin-compaction, @aparte/plugin-diff, @aparte/plugin-echarts, @aparte/plugin-katex, @aparte/plugin-marked, @aparte/plugin-mermaid, @aparte/plugin-model-selector, @aparte/plugin-shiki, @aparte/plugin-storage-idb, @aparte/plugin-streaming-markdown, @aparte/plugin-titler, @aparte/plugin-triggers, @aparte/angular, @aparte/react, @aparte/svelte, @aparte/vue, @aparte/docs-mcp.