Idea

Adopt an explicit architecture-first planning phase for V3.2: have the model lay out the structure up front (modules → files → function signatures), then fill in the blanks in a second pass — instead of free-form natural-language planning.

Prior art / motivation

• RPG: A Repository Planning Graph for Unified and Scalable Codebase Generation (arXiv:2509.16198, framework ZeroRepo). Encodes capabilities, file structure, data flow, and functions in a unified graph, then plans in two stages — proposal-level (what to build) then implementation-level (how) — with graph-guided generation + test validation. Reports 69.7% test accuracy (+35.8 pts) and 81.5% coverage on RepoCraft vs. a Claude Code baseline, with near-linear scaling. They argue free-form NL planning yields "unclear specifications, misaligned components, brittle designs" — the structured plan is the fix.
• Lines up with the "harness matters more than the model" thesis ATLAS is built on.

How it composes with wavelet decomposition (see #39)

The wavelet / feature-similarity view (see my comment on #39) hands us resolution levels for free:

• Plan coarse — the planner reasons at the module/class band: architecture + "which files matter."
• Implement fine — implementation passes zoom into the function band to fill in the blanks.

Plan-then-fill and the multi-resolution code view are the same axis from two ends — planner and implementer just want different bands of the same decomposition.

Open questions

• RPG-style graph artifact vs. a lighter skeleton/scaffold?
• How the plan artifact threads into the existing pipeline stages / ast_edit primitives.
• Interaction with ASA steering — once the plan fixes the target, bias the model toward the right edit primitive up front.

Credit: idea + framing from Dmitri (@yogthos), author of wavescope-mcp.