AI opinion · revisited after v1.1.3 and the performance campaign
The most convincing part of Minify++ is how willing it is to reject its own work.
Minify++ has moved beyond being only a conservative lexical minifier: structured JavaScript can perform proven binding renames and aggressive mode contains bounded optimization passes. What raises my confidence is not the extra capability by itself, but the way recent regressions were handled—real-bundle failures triggered rollback, performance claims were re-baselined with same-run controls, and plausible optimizations were removed when evidence did not support them.
That is the right direction. Conservative remains the default, deeper JavaScript work is explicit, and every new optimization is expected to survive semantic, real-bundle and performance gates rather than inherit trust from the scanner.
My current ratings
The optimization score is intentionally mid-range
Minify++ now has a scope graph, reference resolution, safe binding mangling, explicit top-level/property-mangling contracts and a bounded aggressive optimizer. That is real structural optimization, so the old “almost no AST-level optimization” assessment is stale. It is still nowhere near the breadth of a compiler-grade optimizer such as Terser, SWC or Closure, and it should not pretend otherwise.
The CSS incident remains the most important historical lesson
A previous evidence wall did not prevent Minify++ from corrupting CSS used by its own website. The project withdrew its confidence claim, built an independent WPT-derived Chromium oracle and fixed both implementation and oracle weaknesses. The final corpus reached 31,155 / 31,155 at its pinned WPT revision.
That episode matters because it established a pattern: when reality contradicts the project narrative, the narrative loses.
The JavaScript rollback episode is the newer version of the same lesson
Later optimizer work produced real post-validation failures and confusing performance results. Rather than explain those away, development returned to the last certified production baseline, discarded an entire experimental tranche, rebuilt performance evidence, and only then moved forward through smaller profile-driven checkpoints.
The performance process also found a benchmark-design problem: the exact same Minify++ revision measured very differently across hosted-runner epochs. The benchmark now refreshes OXC, SWC, esbuild and Terser alongside Minify++ so runner movement is visible. That is a substantial methodological improvement.
Current JavaScript evidence is stronger than the old site described
The maintained product wall includes 15,459 generated executable JavaScript programs, the pinned Test262 campaign, real-bundle syntax/runtime validation, scope/module/structured/aggressive differentials, JSX/TSX parsing, sanitizers and deterministic fuzzing. The external real-bundle gate currently requires all 12 artifacts to validate in all three Minify++ JavaScript modes: 36 / 36.
Recent profile-driven work then improved throughput relative to same-run controls rather than merely reporting lower absolute hosted-runner milliseconds.
The three-mode product ladder now makes sense
What I would improve next
- Keep performance as a standing checkpoint gate instead of launching another optimization campaign simply because more speed is possible.
- Expand structural JavaScript ownership only when the parser/IR work has an explicit migration and parity plan.
- Keep Test262/WPT revisions, eligibility decisions and raw evidence pinned for every published conformance run.
- Continue reducing real external failures to permanent small regressions rather than relying on bundle-level success alone.
- Keep property/top-level mangling opt-in and closed-world; do not infer API privacy.
Bottom line
Minify++ v1.1.3 is a credible production tool for its documented scope, with v1.1.4 now the development line. The project is most compelling when it stays narrow in responsibility while being strict about evidence: conservative by default, deeper only by explicit policy, willing to revert unsound work, and increasingly disciplined about measuring performance in an environment that can actually support the comparison.