feat: accurate risk scoring — differentiate Added/Modified/Deleted in detect_changes

[marxo126]

Summary

detect_changes previously marked all symbols in changed files as "Modified", making risk scores meaningless for feature branches. A branch adding 300 new widgets and intents would score HIGH risk — the same as a branch that deletes 50 core functions.

This PR fixes that by:

• Classifying each symbol as Added, Modified, or Deleted using tree-sitter diffing
• Weighting risk by actual danger: score = (modified_with_callers × 3) + (deleted × 5)
• Adding a human-readable explanation field to the summary

How it works

1. Added files (git diff --diff-filter=A): all symbols → "Added" (zero parsing needed)
2. Modified files (git diff --diff-filter=M): parse the old version with tree-sitter via git show base_ref:file, extract symbol names, diff against current index → each symbol classified as "Added" (new in this file) or "Modified" (existed before)
3. Deleted files (git diff --diff-filter=D): parse old version to identify removed symbols → "Deleted"
4. Risk formula: only modified symbols with upstream callers and deleted symbols contribute to risk. New code cannot break existing code.

Real-world results

PricePal v1.3 (iOS 26 app — added widgets, intents, services)

	Before (broken)	After (this PR)
Added	0	514
Modified	711 (everything)	188
Deleted	0	9
Risk	HIGH (meaningless)	HIGH (57) — justified
Explanation	(none)	"514 new symbols added (low risk — no existing code depends on them yet). 188 symbols modified, affecting 3 execution flows. 9 symbols removed — any code that calls them will break."

CuePrompt PiP rewrite (macOS/iOS teleprompter — heavy refactor)

	Before (broken)	After (this PR)
Added	0	2
Modified	120 (everything)	73
Deleted	0	45
Risk	some process count	CRITICAL (261) — correct
Explanation	(none)	"2 new symbols added (low risk). 73 symbols modified, affecting 26 execution flows. 45 symbols removed — any code that calls them will break."

PricePal scored HIGH because it's mostly new code with a few core files touched. CuePrompt scored CRITICAL because 45 symbols were deleted during a PiP rewrite — that's genuinely dangerous. The old tool would have given similar meaningless scores for both.

Changes

src/mcp/local/local-backend.ts

• extractSymbolNamesFromContent() — new helper: parses file content with tree-sitter and returns Set<"Type:name"> for diffing
• detectChanges() — rewritten:
  • Uses --diff-filter=A/M/D instead of --name-only
  • For modified files: parses old version at base_ref to diff symbol sets
  • For deleted files: parses old version to identify removed symbols
  • Extracts node type from ID prefix (LadybugDB doesn't support labels())
  • Weighted risk formula instead of simple process count
  • Human-readable explanation in summary
  • Skips process lookup for Added symbols (optimization — no callers possible)

src/mcp/tools.ts

• Updated tool description to document change types and risk formula

New summary fields

{
  "changed_count": 711,
  "added_count": 514,
  "modified_count": 188,
  "deleted_count": 9,
  "affected_count": 3,
  "changed_files": 39,
  "risk_level": "high",
  "risk_score": 57,
  "explanation": "514 new symbols added (low risk — no existing code depends on them yet). 188 symbols modified, affecting 3 execution flows. 9 symbols removed — any code that calls them will break."
}

Backward compatibility

• changed_symbols[].change_type was always "Modified" before — now it can be "Added", "Modified", or "Deleted". Consumers that only checked for "Modified" will still work.
• New summary fields (added_count, modified_count, deleted_count, risk_score, explanation) are additive — existing consumers won't break.
• risk_level values unchanged: low, medium, high, critical.

Test plan

• Tested on PricePal (61-file iOS 26 app, 39 changed files, 711 symbols)
• Tested on CuePrompt (19-file macOS/iOS app, 4 changed files, 120 symbols)
• All 3,589 existing tests pass (2 pre-existing Swift return-type failures unrelated)
• TypeScript compiles clean
• Works across all scopes: compare, staged, all, unstaged

Fixes #415

🤖 Generated with Claude Code

[vercel]

@marxo126 is attempting to deploy a commit to the NexusCore Team on Vercel.

A member of the Team first needs to authorize it.