Dependency-Graph-Guided Planning

Maintain a repository dependency graph to formulate coordinated multi-file edit plans across complex software projects.

When autonomous coding agents modify large repositories, planning edits strictly through conversational context or sequential tool calls leads to hallucinations of missing imports, broken cross-module call sites, and partial refactorings.

Dependency-graph-guided planning (exemplified by systems like CodePlan) maintains a global dependency graph—tracking abstract syntax trees, call graphs, type references, and file imports across the entire codebase—to formulate coordinated multi-file edit plans:

  1. Impact analysis: Pinpoints which downstream files, modules, and test suites are affected by a proposed change before any file edit is executed.
  2. Topological edit scheduling: Orders code generation and file modifications so that upstream interfaces and data contracts are updated before dependent downstream implementations.
  3. Targeted context pruning: Extracts only the directly impacted definitions and call sites into the agent's context budget, preserving context engineering limits under lost in the middle.

By grounding the agent's plan-validate-execute loop in static dependency structures, this strategy enables multi-file changes without requiring full-repository context dumps or suffering from broken multi-hop dependencies.

Sources

  • Tang & Runkler (2024), sec. 3.2