53 lines
No EOL
2.8 KiB
Markdown
53 lines
No EOL
2.8 KiB
Markdown
---
|
|
sidebar_label: 'Power Steering'
|
|
---
|
|
import Codicon from '@site/src/components/Codicon';
|
|
|
|
# Power Steering (Experimental Feature)
|
|
|
|
The "Power Steering" experimental feature (`POWER_STEERING`) is designed to enhance the consistency of Roo Code's responses by more frequently reminding the underlying Large Language Model (LLM) about its current mode definition and any custom instructions.
|
|
|
|
## How It Works
|
|
|
|
When Power Steering is enabled, Roo Code constantly reinforces the LLM's understanding of its assigned role (e.g., "You are a helpful coding assistant") and any specific guidelines provided by the user (e.g., "Always provide code examples in Python").
|
|
|
|
This is achieved by explicitly including the `modeDetails.roleDefinition` and `modeDetails.customInstructions` within the information sent to the LLM with each interaction.
|
|
|
|
**Goal:**
|
|
The primary goal is to ensure the LLM adheres more strictly to its defined persona and follows user-specific instructions more consistently. If you find Roo deviating from its role or overlooking custom rules, Power Steering can help maintain its focus.
|
|
|
|
**Trade-off:**
|
|
These frequent reminders consume additional tokens in each message sent to the LLM. This means:
|
|
* Increased token usage per message.
|
|
* Potentially higher operational costs.
|
|
* The context window may be filled more quickly.
|
|
|
|
It's a balance between stricter adherence to instructions and resource consumption.
|
|
|
|
**Default Status:** Disabled.
|
|
|
|
## Technical Details
|
|
|
|
* **Experiment ID:** `powerSteering`
|
|
* **Mechanism:**
|
|
* The feature's status is checked by the `getEnvironmentDetails` function.
|
|
* If enabled, the current mode's `roleDefinition` and `customInstructions` are added to the details sent to the LLM.
|
|
* These details are wrapped in `<environment_details>` tags and become part of the context for each LLM interaction.
|
|
* **Impact:** By frequently including the role definition and custom instructions, the LLM is steered to generate responses more aligned with these parameters.
|
|
|
|
## Enabling This Feature
|
|
|
|
Power Steering is managed within the "Experimental Features" section of Roo Code's Advanced Settings.
|
|
|
|
1. Open Roo Code settings (<Codicon name="gear" /> icon in the top right corner).
|
|
2. Navigate to "Advanced Settings".
|
|
3. Locate the "Experimental Features" area.
|
|
4. Toggle the "Power Steering" option.
|
|
5. Save your changes.
|
|
<img src="/img/power-steering/power-steering.png" alt="Settings for Intelligent Context Condensation and Power Steering" width="600" />
|
|
|
|
For general information on experimental features, see [Experimental Features Overview](/features/experimental/experimental-features).
|
|
|
|
## Feedback
|
|
|
|
Please report any issues or suggestions regarding this feature on the [Roo Code GitHub Issues page](https://github.com/RooVetGit/Roo-Code/issues). Your feedback is crucial for improving Roo Code. |