You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
New crate `rpg-build` produces an RPGraph from a natural-language project
specification, inverse of rpg-encoder which extracts a graph from existing
source code. Inspired by the ZeroRepo paper (Luo et al., 2026,
arXiv:2509.16198), limited to the design phases — code generation is
left to the connected coding agent.
Architecture:
- `rpg-build::design_rpg(spec, config)` returns `(RPGraph, DesignReport)`
- Reuses rpg-lift's LlmProvider abstraction (Anthropic / OpenAI /
OpenRouter / Gemini)
- LLM produces a strict JSON schema (3-level hierarchy, entities with
features, calls, imports)
- Parser extracts the JSON from fenced code blocks or balanced braces,
validates types, drops dangling dependency references silently
- Output graph has full hierarchy + edges + features but no source code,
no line ranges (line_start/end = 0)
New MCP tool `design_rpg`:
- Same provider/api_key/api_key_env pattern as auto_lift
- Returns the design rendered as a semantic_snapshot
- Optional `save=true` writes to .rpg/graph.json (clears auto-sync
markers because the new graph has no committed correspondence yet)
- Feature-gated behind `rpg-build` (default ON, --no-default-features
builds without it)
8 unit tests for the design parser + graph builder. All 660 workspace
tests pass.
Bumps to v0.9.0 (minor: new public crate added to the workspace).
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
0 commit comments