🐛 Problem Statement
When using write-plan on medium-to-large codebases, the current context retrieval strategy relies heavily on iterative grep, glob, and read tool calls to build a mental model of the codebase. This results in:
Excessive token consumption: Planning phase alone often consumes 50k–200k+ tokens due to blind scanning.
Long latency: write-plan can take 3–10 minutes before producing any output.
Unreliable plans: The model sometimes misses critical dependencies because it hit context limits during exploration.
I have manually configured CodeGraph MCP + CLAUDE.md instructions as a workaround, which reduced planning tool calls by ~90% and tokens by ~70%. However, this requires manual per-project setup and is not integrated into Superpowers' core workflow.
💡 Proposed Solution
Integrate CodeGraph (or a similar local code-graph indexing mechanism) as an optional but recommended code-map provider for write-plan:
Pre-indexing step: Add a superpowers index or auto-detect command that builds a local semantic graph of the codebase before planning.
Native integration in write-plan: When CodeGraph is available, automatically prioritize structured graph queries (query_graph, trace_call_path) over raw file scanning during the exploration phase.
Fallback mechanism: Gracefully fall back to current grep/glob behavior when CodeGraph is not installed or indexed.
CLAUDE.md auto-generation: Optionally inject CodeGraph collaboration protocol into CLAUDE.md during project initialization.
🐛 Problem Statement
When using write-plan on medium-to-large codebases, the current context retrieval strategy relies heavily on iterative grep, glob, and read tool calls to build a mental model of the codebase. This results in:
Excessive token consumption: Planning phase alone often consumes 50k–200k+ tokens due to blind scanning.
Long latency: write-plan can take 3–10 minutes before producing any output.
Unreliable plans: The model sometimes misses critical dependencies because it hit context limits during exploration.
I have manually configured CodeGraph MCP + CLAUDE.md instructions as a workaround, which reduced planning tool calls by ~90% and tokens by ~70%. However, this requires manual per-project setup and is not integrated into Superpowers' core workflow.
💡 Proposed Solution
Integrate CodeGraph (or a similar local code-graph indexing mechanism) as an optional but recommended code-map provider for write-plan:
Pre-indexing step: Add a superpowers index or auto-detect command that builds a local semantic graph of the codebase before planning.
Native integration in write-plan: When CodeGraph is available, automatically prioritize structured graph queries (query_graph, trace_call_path) over raw file scanning during the exploration phase.
Fallback mechanism: Gracefully fall back to current grep/glob behavior when CodeGraph is not installed or indexed.
CLAUDE.md auto-generation: Optionally inject CodeGraph collaboration protocol into CLAUDE.md during project initialization.