Repository navigation
feat(context): adaptive depth + centrality budgeting + L3 domain-graph for cgis context #220
Description
Follow-ups for the GraphRAG prompt compiler (#19, shipped in #219), capturing design extensions raised by @kehansama on #19. v1 ships a fixed --depth 2, INTERNAL-only neighbours, source snippet, and a lightweight <domain> block; the items below are the next rungs that are specific to the cgis context compiler.
Note: @kehansama's broader insights already have homes — #25 (Prompt-as-Memory / multi-dimensional change memory + negative examples), #16 (DuckDB analytical layer / centrality), #60 (hybrid vector-graph retrieval). This issue intentionally scopes ONLY the compiler-side work and points at those rather than duplicating them.
1. Adaptive depth (leaf vs hub)
Replace the fixed depth with a strategy keyed on the focal node's position in the graph: a bug-fix in a leaf wants depth 1 (immediate neighbours); a refactor of a core abstraction wants depth 3+. @kehansama's heuristic:
depth = max(1, min(5, round(log(out_degree + 1))))
Cheap — one out-degree query before traversal. Self-contained to context_service.
2. Centrality-ranked token budgeting (consumes #16)
For hub nodes the subgraph can still explode. Rank subgraph nodes by betweenness/degree centrality and prune low-centrality branches when the estimated token count crosses a budget — keeping structurally-important context, dropping noise. The centrality computation is exactly what #16 (DuckDB layer) provides; this issue is just the budget-pruning consumer in the compiler. Blocked on #16.
3. Full L3 domain-graph in the <domain> block
v1 surfaces only the focal node's own domain tags. Extend <domain> to the domain-to-domain dependency sub-graph ("this function is part of the auth boundary; see also the permission model") so semantic tasks ("add authentication to this API") get why it matters, not just what connects to what. Builds on the quotient/#186 domain-graph + #47 ontology work.
4. Optional --format mermaid
Reuse MermaidCompiler for a human/visual rendering of the same subgraph, behind a flag — keeping the lean XML lists the default.
Credit: insights from @kehansama's review on #19. Meta-memory / RAG-eval (TruLens, context-effectiveness scoring, negative examples) tracked on #25, not here.
Activity
feat(metrics): DuckDB analytical layer — cgis metrics (#16, slice 1)
feat(metrics): DuckDB analytical layer — `cgis metrics` (#16, slice 1) (
zaebee commentedon Jun 13, 2026
Note for the centrality-budgeting item: betweenness centrality lands here (reassigned from #231). Rationale: betweenness is all-pairs-shortest-paths, impractical as a whole-graph SQL query — but the pruning operates on the small ego-graph the PromptCompiler (#19) already builds, so a plain-Python Brandes' over that subgraph is cheap and exact. The DuckDB layer (#16/#233) provides whole-graph PageRank for the cheaper 'rank by importance' variant if a SQL-side approximation is preferred.
feat(metrics): vectorized PageRank — critical-nodes ranking (#231) (#233
zaebee commentedon Oct 2, 2026
Absorbing #60 (hybrid vector-graph retrieval) as the semantic rung after the lexical one (#206), and the context-compression measurement from #36 (CGIS context bytes vs. naive file dump).
Generated by Claude Code
zaebee commentedon Oct 2, 2026
Prior art for "centrality budgeting", from colbymchenry/codegraph (MIT). Its
codegraph_exploreis the closest thing to what this issue describes (src/mcp/tools.ts):cgis contextneeds a known FQN, so an agent has tofindfirst.computeGraphRelevance, ~L3998). A file with a text match but no graph mass is dropped. The 1–2 most central matching files get a bigger budget.getExploreOutputBudget, L287–380): 13K chars below 150 files, up to 24K above 500 (kept under the host's ~25K inline tool-result limit). It is split across files in proportion to relevance, with at most 70% per file. Siblings off the main path are reduced to skeletons.Suggested split: keep this issue for adaptive depth plus RWR-based budgeting, add an optional free-text
queryentry point, and measure with the agent A/B in #543 (their "sufficiency" metric shows whether the budget went to the right files).Notes:
research/competitor-codegraph.md(idea 1).Generated by Claude Code
feat(bench): agent A/B harness for cgis vs no cgis (#547)