ucn 3.2.0 → 3.3.0

package/test/reliability-test-prompt.md

@@ -0,0 +1,58 @@
+Run UCN reliability tests on 5 real-world projects. IMPORTANT: use node /Users/mihail/ucn/cli/index.js (NOT the global ucn) to test the current dev version. Use --clear-cache on every command.
+
+First run the unit tests: node --test test/parser.test.js — should be 212 pass, 0 fail.
+
+1. Clone these repos to /tmp/ucn-reliability-test-2/ (shallow clone --depth 1):
+   - fastify: https://github.com/fastify/fastify.git (JavaScript)
+   - httpx: https://github.com/encode/httpx.git (Python)
+   - hugo: https://github.com/gohugoio/hugo.git (Go)
+   - alacritty: https://github.com/alacritty/alacritty.git (Rust)
+   - spring-boot: https://github.com/spring-projects/spring-boot.git (Java — use spring-boot-project/spring-boot/src/main/java/org/springframework/boot only, pass that as the target path)
+
+2. Launch 5 parallel agents (one per project). Each agent should run ALL of these commands with --clear-cache and 30s timeout:
+
+   toc, toc --detailed, find <sym1>, find <sym2>, about <sym1>,
+   context <class>, context <function>, impact <sym1>, smart <sym1>,
+   trace <sym1> --depth=2, usages <sym2>, deadcode, fn <method>,
+   imports <file>, exporters <file>, search "<term>", verify <sym1>,
+   find nonexistentXYZ (edge case), context "" (edge case)
+
+3. Key symbols per project:
+   - fastify: route, FastifyInstance, inject, Reply, register
+   - httpx: get, Client, request, Response, __init__
+   - hugo: Build, Site, Execute, Page, New
+   - alacritty: main, update, Display, build, new
+   - spring-boot: run, SpringApplication, main, ApplicationContext, refresh
+
+   For `fn` command, specifically test class methods (Python __init__, Java overloaded methods) — these were broken before and fixed in this version.
+
+4. Focus areas for EVERY command — check ALL of the following (AST tree-sitter reliability is the main focus):
+   - toc shows correct file/function/class counts, no crashes on large projects
+   - toc --detailed lists all symbols with line ranges, no truncation errors
+   - find returns ranked results by usage count, no duplicates
+   - about prefers lib/src/core definitions over test/example files, shows usages/callers/callees/code
+   - context <class> shows class name (not "undefined"), lists methods with signatures
+   - context <function> shows callers and callees with correct counts and weights
+   - context for non-existent symbols returns "Symbol not found" (NOT empty callers/callees)
+   - impact groups call sites by file with code context, shows argument patterns
+   - smart shows function code + inlined dependency code, overloads show ALL overload callees
+   - trace builds correct call tree with depth levels, weights, and multiplicity
+   - usages categorizes by type (definition/call/import/reference), no false positives
+   - deadcode completes within timeout, correctly excludes entry points (main, init, __init__, #[test])
+   - fn extracts correct code for BOTH top-level functions AND class methods (previously broken, now fixed)
+   - fn auto-resolves best definition (prefers lib/src over test)
+   - imports shows internal + external imports with resolved paths
+   - exporters finds importers for the file (Python relative imports work, Java package imports resolve)
+   - search returns matches with file grouping, respects project ignores
+   - verify checks call signatures, totalCalls equals valid+mismatches+uncertain (no inflated counts from filtered method calls)
+   - Edge cases: find nonexistentXYZ returns clean "not found", context "" returns usage error (no crash)
+
+5. Each agent reports: command, status (OK/ERROR/CRASH/UNEXPECTED), notes.
+   At the end: total pass/fail, issues by severity, patterns.
+
+6. After all agents finish, compile a cross-project summary comparing results to the previous run (which had 85/95 = 89.5% before fixes, target is 95%+ with 0 crashes). The fixes applied were:
+   - fn command now uses symbol index startLine/endLine directly (fixes class method extraction)
+   - verify totalCalls now computed as valid+mismatches+uncertain (fixes inflated counts)
+   - context returns null for undefined symbols (CLI shows "Symbol not found")
+
+All code uses tree-sitter AST parsing — reliability and correctness of AST-based analysis is the main focus. Any crash, incorrect result, or hang is a bug.