본문으로 건너뛰기

Managing Inputs

FluxLoop manages inputs through the Inputs view in the VSCode extension. Inputs are the prompts and data your AI agent processes during experiments.

Viewing Inputs

Open the Inputs view in the FluxLoop activity bar to see everything related to scenarios:

  • Configuration – Settings and base input from configs/input.yaml
  • Generated Inputs – Variations created via the wizard, annotated with strategy + persona

Clicking any entry opens the underlying YAML. The panel refreshes automatically whenever configuration files change.

Configuration

The Configuration folder contains:

Open Configuration

Click to open configs/input.yaml directly in the editor.

Base Input

Shows the primary seed input (first entry in base_inputs) that drives generation:

base_inputs:
- input: "How can I track my order?"
expected_intent: "order_tracking"

The base input is the foundation for all variation strategies. Only the first base input is displayed in the tree for simplicity.

Generating Input Variations

Generate variations from base inputs using LLM or deterministic strategies.

From Inputs View

  1. Click Generate New Inputs…
  2. Configure the wizard:
    • Generation Mode – Deterministic or LLM
    • Base Input Confirmation – Review the primary base input
      • Select Use Input to proceed
      • Select Edit in configuration to open configs/input.yaml and modify
    • Strategies – Pick one or more strategies (rephrase, verbose, concise, error_prone, etc.)
      • Pre-selected based on variation_strategies in configs/input.yaml
      • Your selections are saved back to variation_strategies for next time
    • Variation Limit – Number of outputs per base input
    • Output File – Defaults to inputs/generated.yaml (auto-populated from config)
    • Overwrite / Append – Control whether the existing file should be replaced
    • Dry Run – Preview results without writing files
    • LLM Provider & API Key – Prompt appears for LLM mode; the key can be saved to VS Code secret storage and appended to .env if you approve

Generated inputs immediately appear under Generated Inputs with metadata (strategy, persona, timestamp). Double-click any entry to open the YAML.

From Command Palette

FluxLoop: Generate Inputs

From Terminal

fluxloop generate inputs --mode llm --strategy rephrase --limit 5

Input Strategies

Available variation strategies:

StrategyDescriptionMode
rephraseReword while preserving intentLLM
verboseAdd detail and contextLLM
conciseShorten to essential pointsLLM
error_proneIntroduce typos or ambiguityLLM
typoAdd typos/errorsLLM
persona_basedBased on persona characteristicsLLM
adversarialEdge cases and attacksLLM
multilingualDifferent languagesLLM
customCustom variation promptLLM

Configure default strategies in configs/input.yaml:

variation_strategies:
- rephrase
- verbose

These will be pre-selected in the wizard and updated when you make changes.

Recordings (Advanced)

Recordings capture actual function arguments during runtime for replay in experiments.

Enable Recording Mode

From the Experiments view, expand Recording (Advanced) and click:

  • Enable Recording to turn on argument capture
  • Disable Recording to turn it off

Or from terminal:

fluxloop record enable
fluxloop record disable

This updates:

  • .env: Sets FLUXLOOP_RECORD_ARGS=true/false
  • configs/simulation.yaml: Enables/disables replay_args.enabled

Capture Arguments

  1. Run your application normally
  2. FluxLoop SDK captures arguments to recordings/args_recording.jsonl
  3. Each entry includes function name, args, kwargs, timestamp

Using Recorded Arguments in Experiments

Edit configs/simulation.yaml:

replay_args:
enabled: true
recording_file: "recordings/args_recording.jsonl"
override_param_path: "data.content" # Path to inject generated input

Now when you run experiments, FluxLoop:

  1. Loads recorded arguments
  2. Replaces the field at override_param_path with generated input variations
  3. Replays each modified call

Environment Considerations

Input generation and experiments depend on your Python environment.

Ensure Correct Environment

Before generating inputs or running experiments:

  1. Run FluxLoop: Show Environment Info
  2. Confirm CLI, SDK, and MCP packages live inside the detected interpreter
  3. Adjust with FluxLoop: Select Environment (Auto / Workspace / Global / Custom)
  4. If you run commands with wrappers (uv run, pipx run), configure them once via FluxLoop: Configure Execution Wrapper

See Environment Configuration for the full matrix.

Next Steps