DevDiff

Appearance

How It Works ​

DevDiff sits between your Git history and your collaboration tools, serving as an intelligent processing and sanitization layer.

mermaid
flowchart TD
    A[Git Commit / Hook / CLI trigger] --> B[AST Parser & Trimmer]
    B -->|Prunes unmodified code| C[Injection Guard V2]
    C -->|Detects prompt & shell injections| D[Secret Redactor V2]
    D -->|Redacts API keys, passwords, JWTs| E[Progressive Chunking & Dynamic Timeout]
    E -->|Optimizes payload sizes & times| F[Intelligent Router]
    F -->|Local Ollama / Cloud API| G[Persona Engine Formatter]
    G -->|Applies developer, ceo, compliance tone| H[Output Deliverer]
    H -->|Markdown Logs / Mermaid Diagrams / JSON / Slack| I[Output]

Detailed Pipeline Breakdown ​

Step 1: Git Watcher & Event Ingestion ​

The process begins when code changes are made. Ingestion happens in one of three ways:

  1. Local Execution: Triggered directly via devdiff generate in the CLI.
  2. Commit Watcher Daemon: The GitWatcher listens to local git events or filesystem modifications using chokidar.
  3. Webhook Receivers: The DevDiff Gateway exposes endpoints to listen to incoming webhooks from GitHub Actions, GitLab CI, or Linear.

Step 2: AST Trimming & Secrets Redaction ​

Before code changes are sent to any AI model, DevDiff processes the diff to optimize tokens and ensure strict privacy:

  • AST Trimming: Instead of uploading entire source code files, DevDiff parses the code into an Abstract Syntax Tree (AST), identifies the precise functions or blocks that were modified, and discards the rest of the unmodified syntax.
  • Secrets Redaction: An automated scanner matches standard API keys, passwords, and security token signatures (e.g. AWS keys, Slack webhooks). Discovered keys are immediately replaced with [REDACTED] to prevent leakages to cloud endpoints.

Step 3: Token-Aware Intelligent Routing ​

The processed diff is analyzed to compute its size (token estimate) and complexity (e.g., number of files, breaking changes). The Intelligent Router then:

  • Evaluates the capabilities of configured models (e.g., local Ollama vs cloud GPT-4o).
  • Selects the most cost-effective and capable model that fits within context window limits.
  • Organizes fallback chains to gracefully switch models if the primary provider fails.

Step 4: Persona Engine Formatter ​

Once the model produces the raw analytical summary, the Persona Engine intercepts the output. It maps tone, formatting, emoji rules, and verbosity constraints defined in the chosen persona profile:

  • Developer Persona: Highly technical, showing affected files and AST function signatures.
  • Executive/CEO Persona: Brief, high-level business impact focus, avoiding technical jargon.
  • Compliance Persona: Risk-focused, security-audit styled reports.

Step 5: Multi-Format Delivery ​

The formatted response is delivered to configured outputs, supporting Markdown logs, interactive Mermaid diagrams, JSON outputs, or Slack thread messages.