feat(core): AST-based pattern analysis via tree-sitter queries

[prosdev]

Summary

Adds AST-based pattern detection to dev_patterns using tree-sitter queries. Replaces regex-only detection with 12 S-expression queries across 3 categories, covering the full JS/TS ecosystem (.ts, .tsx, .js, .jsx).

What it does

Error handling — detects patterns regex can't:

Pattern	Before (regex)	After (AST)
throw new Error(...)	Yes	Yes
Result<T> / { ok: true }	Yes	Yes
try { } catch { }	No	Yes
promise.catch(handler)	No	Yes
await inside try/catch	No	Yes
class AppError extends BaseError	No	Yes

Import style — more precise detection:

• Dynamic import() — now counts as ESM (was invisible)
• require() — AST-precise (regex was fragile)
• Re-exports — distinguished from regular imports

Type coverage — catches what regex misses:

• Arrow function return types ((): Type => ...) — regex is fragile on these
• Accurate denominator via arrow-total + function-total queries

Architecture

PatternAnalysisService
    │ calls runAllAstQueries() ONCE per file
    ▼
PatternMatcher (interface — 1 method)
    └── WasmPatternMatcher (web-tree-sitter, WASM)
          └── runQueries() — tree.delete() + query.delete() in finally

Extractors are synchronous pure functions that take pre-computed AST map.
Rules are S-expression string constants (no query builder — YAGNI).
Designed for future swap to @ast-grep/napi if bulk scanning perf matters.

Changes

New files

• packages/core/src/pattern-matcher/ — PatternMatcher interface, 12 rules, WASM implementation
• packages/core/src/pattern-matcher/__tests__/wasm-matcher.test.ts — 51 tests

Modified files

• packages/core/src/scanner/tree-sitter.ts — add TS/TSX/JS languages + runQueries()
• packages/dev-agent/scripts/copy-wasm.js — bundle 3 new WASM grammars
• packages/core/src/services/pattern-analysis-service.ts — AST-enhanced extractors + runAllAstQueries
• packages/core/src/services/pattern-analysis-types.ts — optional patternMatcher in config
• packages/mcp-server/ — wire PatternMatcher through InspectAdapter

Bundle size

+5.3MB WASM (TS 2.3MB + TSX 2.3MB + JS 647KB). Acceptable for CLI tool.

Test plan

Automated (51 tests in wasm-matcher.test.ts)

• 10 positive tests with exact match counts (count === 1)
• 10 negative tests (one per query, count === 0)
• 3 total count queries (arrow-total, function-total, no functions)
• 3 language routing tests (TSX fixture, JSX→javascript, unsupported)
• 3 edge cases (empty source, malformed TS, invalid S-expression)
• 1 performance sanity (552 lines + 12 queries in 39ms)
• 5 resolveLanguage extension routing tests
• 7 extractErrorHandlingWithAst merge logic tests
• 4 extractImportStyleWithAst merge logic tests
• 5 extractTypeCoverageWithAst merge logic tests (accurate denominator)
• 49 existing pattern-analysis-service tests pass unchanged (regex fallback)
• 1628 total tests pass, 0 failures

Manual smoke test

pnpm build                          # Bundles 4 WASM grammars (go, ts, tsx, js)
pnpm test -- wasm-matcher           # 51 tests, 710ms ✓
pnpm test -- pattern-analysis       # 49 passed, 1 skipped ✓

# Full MCP flow (requires Antfly running):
dev index
# In Cursor/Claude Code:
# > Use dev_patterns with filePath "packages/core/src/services/pattern-analysis-service.ts"
# > Use dev_patterns with filePath "packages/core/src/services/pattern-analysis-service.ts" format "json"

Known test gap

• InspectAdapter test does not construct with a real PatternMatcher — AST path is not exercised through the MCP adapter layer. Tracked in .claude/scratchpad.md for follow-up.

Review findings addressed

Round	Finding	Fix
Code review	Query objects not deleted (WASM leak)	q.delete() in finally
Code review	Type coverage ratio inflation	Added arrow-total + function-total denominator queries
Code review	try-catch alone classified as 'throw'	Falls through to regex instead
Code review	Import count asymmetry	Documented with comment
Test coverage	Extractors never tested with real AST	16 extractor merge logic tests added
Test coverage	Denominator queries untested	3 tests for arrow-total/function-total
Senior review	Same source parsed 3x	runAllAstQueries parses once, extractors take pre-computed map
Senior review	runQueries doing too much	Extractors are now sync pure functions

Research that informed this

• Studied ast-grep architecture (custom matcher on tree-sitter AST, not S-expression queries)
• Validated all 12 S-expressions against actual tree-sitter-typescript grammar by parsing real snippets
• Three plan reviews (architecture, WASM/AST specialist, SDET) before implementation
• Decided against ast-grep CLI (shell injection, temp files) and NAPI (premature — WASM is 39ms for 552 lines)

Generated with Claude Code