🚩 2.8.0

Released August 29, 2026

Studio connects git repositories that publish ready-made generators and installs one of them as a project, so a new data source no longer starts from an empty configuration. Projects export and import as ZIP archives, the log is split into a file per component, and every running instance reports the processor, threads, disk, network and queue memory it occupies.

🚀 New features

Eventum Studio

• Generator repositories — connect a git repository of ready-made generators and install any entry of its catalog as a project
• Repository discovery — find a repository by searching GitHub for the ones carrying the eventum-generators topic, and connect it from the list
• Project export and import — export a project as a ZIP archive and import it as a new project, on the same instance or another one
• Monitoring rebuilt around the running instances — the load chart and the instance table share one search, one set of filters and one selection
• Instance overview — an instance opens on its rates, its totals and what the run costs, which used to be a grid of figures at the foot of the page
• Release highlights — what an upgrade brought opens by itself on the first load after it, and stays reachable from the user menu
• Log channels on the Management page — stream the Main, Server, Access and MCP logs of the application separately
• Download in the project file tree — save a project file to your machine, including generator output too large for the editor to open
• A keyring picker beside password fields — pick a secret from the keyring instead of typing the ${secrets.<name>} reference by hand
• Scripts in the scenario data flow diagram — the diagram carries the global keys a script event plugin reads and writes, beside the templates that do the same
• A rebuilt About dialog — the version, the runtime, the host and the build read as one aligned list
• Error details ordered by diagnostic value — the dialog leads with what the server reported, and the raw request and response sit behind one toggle

Core

• A log file per component — a generator writes a log file of its own, down to the traffic its output plugins produce, instead of sharing one with the application
• Per-instance resource accounting — an instance reports the threads, processor time, waiting, disk and network bytes and queue levels it occupies, and Monitoring ranks the running ones by each of them
• generation.queue.max_event_bytes — limit the events queue by the memory it holds, not only by the number of batches in it
• log.third_party_level — set the level of third-party libraries separately from log.level, warning by default
• The state of the GIL — the instance information reports whether generators run in parallel, and Management warns when the GIL came back after startup
• The caller of a served request — everything logged while serving a request carries the address of the client that caused it

Generators

• Global state in the script event plugin — a script reads and writes the same global state a template does, through a globals key of its parameters
• concurrency in the http output — cap how many requests the plugin performs at a time, 100 by default

AI agents

• Repository tools — an agent lists the connected repositories, reads the catalog one publishes and installs a generator from it
• export_generator and import_generator — an agent moves a whole project in or out as a ZIP archive instead of copying it file by file
• get_instance_logs — an agent reads any log channel of the application, not just the log of a generator
• Resources in get_generator_stats — the tool carries the same per-instance resource figures the API does

API

• Repository, export and import endpoints — connect repositories and export or import projects over HTTP

🐛 Bug fixes

Eventum Studio

• A page whose code or stylesheet never arrived reloads the document once, instead of dropping Studio on the error screen
• The Studio shell is no longer served from the browser cache, so it cannot outlive an upgrade and ask for files the new build does not carry
• Wide tables scroll inside their own panel on a narrow window, instead of putting their rightmost columns out of reach behind the page scrollbar
• The editor keeps its width while the explorer and the inspector give way, instead of absorbing every bit of narrowing
• The plugin parameters form follows the plugin rather than its position in the list, so deleting one no longer writes its values over another
• Switching off the events-queue memory limit, or batching by batch.size or batch.delay alone, now reaches the instance: the settings form dropped the value that asks for it, so the choice was never applied and the switch came back on
• The instance settings are reread after saving, so the page stops showing the values it held before the write
• An instance registered from outside the workspace names the directory it sits in, instead of linking to a project page that cannot exist
• The projects table no longer re-renders without end when it is opened without an instance filter
• The template editor completes pop among the state calls it offers
• The table sort controls, the instance switches and the delimiter shortcuts carry names a screen reader can announce

Core and API

• A secret named keys, items or aws.get no longer resolves to a method of the collection holding the secrets, for ${params.*} as well
• What uses a secret counts connected repositories as well as projects, and renaming it repoints every one of them
• Non-ASCII values in a configuration are stored as they were entered, not as \uXXXX escape sequences
• Configurations, templates, samples, time patterns and log files are read and written as UTF-8 whatever the host locale is
• The pipeline backpressure warnings name what actually happened — a cancelled write, or a queue that filled up — instead of blaming the generation rate
• Repeated write timeouts of an output plugin are reported as one line with a running count, in place of one line per cancelled write
• A path that leaves the project directory through a symlink is refused for every file operation

Generators

• A tcp output whose target stopped reading fails the write at once, instead of growing the memory of the application and holding up the shutdown of the instance
• A template that acquired the global state lock and never released it no longer freezes every other generator
• A batch written through the http output under a per-event formatter no longer fires a request per event at once, which used to cancel the write on its timeout and count every event as failed
• A subprocess.run call runs under a timeout of 30 seconds by default and 300 at most, and fails rather than holding more than 8 MiB of output

⚡ Performance

• A generator catching up — started on a past range, or fallen behind real time — merges the batches whose timestamps are already due up to batch.size, instead of publishing each of the small batches batch.delay cut them into
• Starting many instances at once no longer compiles all of their templates at the same moment, so a large scenario comes up without the spike on the host

📝 Other changes

• The body of POST /preview/{name}/event-plugin/produce is described by a request model of its own and rejects anything else
• The preview global state endpoints moved from /preview/{name}/event-plugin/template/state/global to /preview/{name}/event-plugin/state/global, since the state they read belongs to every event plugin
• The disk figures of the application count the bytes handed to the system calls, the same measure an instance is reported by, so the two can be compared; the application used to count what reached the disk instead