Deterministic Dual-Agent Task Decomposition & Execution

The Planner-Synthesizer Split

Single-agent LLM systems frequently lose context when tasked with multi-file edits. CodeXa bifurcates the cognitive process into two discrete agents:

1. Astra 6 Planner: Analyzes user intent against the AST dependency graph and generates a deterministic DAG (Directed Acyclic Graph) of edit operations.

2. Synthesizer: Executes individual hunk rewrites in isolated memory sandboxes, passing results to Fabel 5.1 for invariant verification.

// Dual-Agent Orchestration Flow
const plan = await Astra6.planTask({
  task: "Refactor auth middleware to handle JWT cookie refresh",
  astGraph: workspace.getSymbolGraph(),
});

for (const step of plan.steps) {
  const diff = await Synthesizer.generateHunk(step);
  const isValid = await Fabel5.verify({ diff, rules: workspace.lintRules });
  if (isValid) {
    workspace.stageHunk(diff);
  }
}

Atomic Hunk Verification

Before any proposed edit is shown to the developer, syntactic AST validation ensures no syntax errors or unclosed brackets exist. CodeXa rejects invalid hunks automatically without interrupting the engineer.

AST-Grounded Prompting

Rather than dumping raw text into the model context window, CodeXa extracts only the relevant type interfaces, exports, and call hierarchies, saving up to 70% of prompt token quotas.