Reference
Global settings
Choose a shared model profile and a data folder for each console task, so you do not repeat those choices. A model profile specifies the provider and model each step calls. You can also save an advice-file selection for later use, but global advice is not yet applied to runs.
Change a setting
- Open Settings and connect. Check the server shown on the page: these folders live on the machine running the console.
- Under Models, choose a profile. Show routes lists each step's provider and model; the page also lists required credential names without showing their values.
- Under Data folder, choose an allowed data root for the task, open subfolders and select Use <folder>. Symbolic links are visible but cannot be selected. Set in console.json means the task already has an explicit data folder.
- Select Save settings. After Saved; active after server restart appears, restart the serve command. Values in effect shows the values the running console uses, their origins and their server paths.
Expected result: after restart, tasks without their own model or data-folder setting use these defaults. Before restart, the page distinguishes saved choices from active values.
Where the files live
From the SIDERIUS checkout, run .venv/bin/python -m tools.candidate_console init --workspace /absolute/project, replacing /absolute/project with your own project directory. This creates siderius_workspace/ under that directory. The parent must be explicit; there is no default location. Running init again only creates missing content and does not overwrite existing files.
settings/console.json holds console configuration, settings/console-token holds the access token, and settings/global_settings.json holds global choices. Model profiles and advice files live in library/llm/ and library/advice/. Initialization saves and selects the standard model profile, named default-<12 hex characters>.json.
data/<task_id>/ holds task data; you create and populate each task folder yourself. operations/, preparations/, candidates/ and runs/ hold console state and run outputs. projects/ is the conventional home for workflow editor projects.
Which choices are available
An allowlist names the files and data locations the browser may choose. The console operator registers them in console.json under llm_profiles, advice_files and data_roots; workspace library files are also available. Global settings save choices by entry ID or library reference, so browser users do not type server paths.
The console reads the settings file only when console.json has a global_settings block. Saving and importing require "global_settings": {"browser_writable": true}. Initialization writes this into new configurations; existing ones need an operator edit followed by check. With false, the console still reads global settings but the browser is read-only.
Which setting wins
Precedence, from lowest to highest, is global settings < console.json < the workflow's own settings. Global choices fill omissions in console.json; a workflow keeps its own explicit values. If neither source chooses a data folder, the workspace layout supplies data/<task_id>/. There is no layout default for a model profile.
For example, selecting global profile A does not replace profile B explicitly set for a task in console.json. A workflow's explicit values still apply to that workflow. Check the origin in Values in effect to understand why A was not used. Advice has a separate limitation: the global selection is stored only and neither replaces nor supplements workflow advice.
Import and export
Under Models or Advice, select Import from this computer… to import a JSON model profile or a command-line-format advice file, including a previous export. The console validates it before copying it to the server library. Importing the same file again reuses its copy; library files are never replaced or deleted. Advice lets you select a file or None, but that selection does not yet affect runs.
Under Export a workflow, select the project, task and export name, then Export. The server writes to siderius_workspace/exported_configs/<export name>/: llm.json is the selected task's active model profile, workflow.py is the last saved workflow, and advice.json appears only when the workflow has advice. Each Download button saves that file to this computer.
Command-line runs read the profile with --llm_config and advice with --advice. Exports exclude credentials, the console token and the credential-file path. They also exclude settings saved but still awaiting restart.