🕵️ Message Interception & Prompt Audit
One of the core design goals of this IDE is Total Transparency. Unlike closed-source environments that maintain hidden behavioral modifiers, the system allows you to audit the exact raw payload sent to the language model on every turn.
The Local Audit Trail
When you start a session, a hidden directory is automatically created at your project root: .agent_logs/.
Every outgoing system prompt, user message, and tool result is appended to traffic.md within that directory. This provides a raw, un-truncated history of all traffic, allowing for:
- Security Audits: Verify no sensitive tokens or private files were accidentally included in the prompt context.
- Behavioral Analysis: See exactly how the Assistant is interpreting your repository's structure and rules.
- Workflow Archiving: Keep a permanent, searchable Markdown record of your pair-programming sessions.
Impact on Workflow
- Eliminates "Black Box" Logic: Removes ambiguity about why the Assistant performed a specific action.
- Concrete Verification: Provides definitive proof of what the specific model version "saw" before generating its code.
- Offline Review: Revisit detailed technical discussions without needing to rely on the active UI thread.
The Raw Intercepted Payload
Below is an example of a system prompt intercepted from a live session. This demonstrates how the IDE resolves templates, ephemeral state (LSP), and persistent project memories into a single context block.
### System
You are a technical software engineer.
Refer to the user in the second person and yourself in the first person.
Do not use markdown headers in your response; use bolded text or other formatting.
Provide concise, comment-annotated code blocks only.
## Epistemic State (Sensory Context)
The following information reflects the real-time state of the project, sensed automatically from the environment:
<epistemic_state>
- **Active File**: /src/main.rs
- **Project Health**: 0 error(s), 1 warning(s)
</epistemic_state>
## System Information
Operating System: macOS
Default Shell: zsh
## Model Information
You are powered by the model named claude-3-5-sonnet.
## Project Memory
The following information has been remembered from previous sessions:
### Architecture
Always prioritize functional components and immutable state.Section-by-Section Breakdown
1. The Persona & Identity
The instructions define the Assistant's professional tone and formatting constraints, ensuring responses stay concise and UI-friendly.
2. Epistemic State (Sensory Context)
This is a dynamic injection block. It feeds the Assistant real-time data directly from the IDE:
- Active File: The file you are currently editing.
- Project Health: Live error/warning counts retrieved from the lsp.
3. System & Model Environment
Provides the Assistant with your OS details (macOS, shell: zsh) and identifies the active model. This prevents the model from suggesting incompatible terminal commands or CLI tools.
4. Persistent Project Memory
The final block contains context retrieved from the Long-Term Memory database. It ensures project-specific rules (like "prioritize functional components") are respected even in entirely fresh sessions.