Aspirational features
These are ideas, not commitments. Each item states why it might matter and what has to be true first.
Segment quality tuning
Find the largest segment that keeps answer quality. Measure at 64K, 128K, and 256K on a fixed question set. The current planner uses the usable context, not a measured quality limit.
Scope configuration
Add a per-repository scope file with include and exclude patterns. Keep the current module as the default. The file should record a reason for each exclusion.
Cross-repo categories
Promote identity, infra, and data from tags to first-class nodes.
A category can span repositories.
The decision log and findings should attach to the category.
Optional embeddings
Add an embedding index as a secondary search path. Keep the deterministic FTS and graph as the baseline. The coverage state should still decide what has been read.
Model-generated labels
Use a model to write a richer label or summary for a chunk. Store the label with the chunk hash and model version. Invalidate it when the chunk changes. The deterministic label remains the fallback.
Multi-agent batch read
Run several segments in parallel with a shared ledger. The ledger needs a lease per segment and a merge step for findings. The workspace is the first candidate.
Editor and IDE integration
Show coverage state in the editor. Open the next unread segment from the project view. The CLI and SQLite ledger are enough for the first version.
Shared or hosted index
Publish a read-only ledger for a repository. A team can share the map and findings. The write path stays local.
Quality benchmark
Build a question set whose answers require obscure chunks. Measure citation support, answer correctness, and meander.
Semantic graph
Use a symbol graph for callers, callees, and tests. The current graph is import-based. The retrieval layer can use the graph as a ranking signal.