Status: stream of thoughts, might be nonsense. Draft for discussion, not a commitment.
Goal: let an agent understand project "quality" — not as an HTML report (that's for humans; an agent won't wait for it), but as structured verdicts through MCP tools: {dead_code: [...], complexity_hotspots: [...], impact: [{node, score, confidence, reason}]}.
Algorithms already studied (from competitors, all small — we write them in-house)
Dead code (CGC, ~40 lines):
- Function with zero incoming CALLS from non-dependency callers
- The real value is in exclusion heuristics:
main/setup/run, dunders, test_*, names containing main/entry/application, decorator filter, is_synthetic for module-level frames
Cyclomatic complexity (CGC, ~20 lines, index-time):
1 + count(node.type ∈ {if, for, while, except, with, boolean_operator, list_comprehension, generator_expression, case_clause})- Stored as a property on Function at indexing time
Impact scoring (CodeSlicer — the best algorithm of the bunch):
chain_confidence = geometric mean of edge confidences along the path (doesn't let a long path collapse the way naive multiplication does)impact_score = criticality(kind) × chain_confidence × decay^distance- Status thresholds: ≥0.90 Confirmed / ≥0.75 High probability / ≥0.55 Needs verification / Ambiguous
Coupling/Cohesion: Ca/Ce, Instability, LCOM/CAM — computed from our own edge graph
Principle
Everything essential — in-house, in one package. Do not shell out to competitor binaries (archscope etc.): the market is winner-take-all; a user will install one tool — if raged depends on someone else's binary, the user will drop raged and install that binary directly.
The foundation already exists: #56 (global symbol table) + confidence on edges.
Relates to: #35 (cross-repo support), #24 (multi-index for third-party libs), #20 (advanced semantic search), #31 (multi-project configs)
Status: stream of thoughts, might be nonsense. Draft for discussion, not a commitment.
Goal: let an agent understand project "quality" — not as an HTML report (that's for humans; an agent won't wait for it), but as structured verdicts through MCP tools:
{dead_code: [...], complexity_hotspots: [...], impact: [{node, score, confidence, reason}]}.Algorithms already studied (from competitors, all small — we write them in-house)
Dead code (CGC, ~40 lines):
main/setup/run, dunders,test_*, names containingmain/entry/application, decorator filter,is_syntheticfor module-level framesCyclomatic complexity (CGC, ~20 lines, index-time):
1 + count(node.type ∈ {if, for, while, except, with, boolean_operator, list_comprehension, generator_expression, case_clause})Impact scoring (CodeSlicer — the best algorithm of the bunch):
chain_confidence= geometric mean of edge confidences along the path (doesn't let a long path collapse the way naive multiplication does)impact_score = criticality(kind) × chain_confidence × decay^distanceCoupling/Cohesion: Ca/Ce, Instability, LCOM/CAM — computed from our own edge graph
Principle
Everything essential — in-house, in one package. Do not shell out to competitor binaries (archscope etc.): the market is winner-take-all; a user will install one tool — if raged depends on someone else's binary, the user will drop raged and install that binary directly.
The foundation already exists: #56 (global symbol table) + confidence on edges.
Relates to: #35 (cross-repo support), #24 (multi-index for third-party libs), #20 (advanced semantic search), #31 (multi-project configs)