PHP Typing Conformance Results

Each row links to a per-test detail page with the source, expectations, and every analyzer’s raw output.

Results last updated: 2026-08-15T00:10:02+09:00

Soundness — potential runtime type errors

Positives here flag code that can actually fail at runtime — type mismatches, null access, invalid arguments, uninitialized reads. Pass means the analyzer agrees with the expected diagnostic; rows whose test marks a declaration with // T instead report recognition and enforcement of the type spelling. See the legend above.

Style & opinionated rules — no runtime-safety impact

Lint-style opinions and advisories (PHPStan strict-rules, deprecations, doc conventions) that do not change whether the code runs. Cells show whether each analyzer opts into reporting the rule — not a pass/fail verdict.

Debug features — type-inspection helpers

Dump/trace APIs used while debugging inference (PHPStan\dumpType(), @psalm-trace, @phan-debug-var, Mago\inspect(), …). Cells use the same recognition/enforcement vocabulary as type-spelling rows: does the analyzer accept this helper, and does it emit a diagnostic that reveals the inferred type?

Analyzers

Reference metadata for each analyzer compared above; versions are those pinned by this suite. Analysis is the one column whose values are coined here rather than read off the project: a type checker aims at type correctness, a type-aware linter at a rule catalogue that inference sharpens, and code intelligence infers types mainly to drive completion and navigation, with diagnostics one feature among many.

Language servers

Every cell in this table records what the project claims about itself — from its README, docs, changelog, release notes, package metadata, or its maintainer’s own blog — and no claim here has been verified by executing the tool. What has been measured lives elsewhere: Intelephense’s and Psalm’s diagnostics scores are in the matrix above, and the servers this machine can launch headless have their protocol behaviour recorded in the measured capability matrix below. Cells read Not stated where the project simply does not say.

Four rows also appear in the analyzer table above. Intelephense and Psalm are in both as the same tool — Psalm’s server runs the analyzer its CLI does, reached over the protocol instead of the command line. The other two are pairs: php-lsp drives mir, and phpy is the CLI frontend of devsense-php-ls. Each pair is versioned separately, so each artifact keeps its own row and its own dates. Phan ships a language server too and belongs here; it has simply not been researched yet.

Language server capabilities — measured

Everything in this section comes from launching each server headless against a small fixture workspace: no cell repeats a claim from a README. Advertised is what the server’s own initialize response declared; the probe verdict says what happened when the suite then called the capability at a position prepared for it. A dash means the handshake did not advertise the capability. The two freemium servers gate paid features in opposite ways, and the matrix shows both faithfully: a free Intelephense omits them from its handshake (they render as dashes), while DEVSENSE advertises them and refuses the call by licence (they render as Gated). Hover any cell for what it does and does not claim.

Hover type conformance

The typing question, asked over the protocol: hovering a variable whose type the annotation or the narrowing has already established, does the server show that type? Precise means the shown type matches the expected spelling; Widened / other means hover answered with something else — typically the declared type where the narrowed one was expected, or a widened fallback — and the cell’s hover note quotes it verbatim.

Navigating a real project’s dependency graph

The corpus is a real project — psy/psysh@0e8ec1af5990 from the steins-survey checkout — and each row is one symbol in it, chosen per kind. Go to definition jumps from a use site and must land on the declaration; find references asks from the declaration and is scored against the complete, grep-verified set of code references, with only the two files involved in each probe opened — everything else must come out of the server’s own index. A +n after the score counts locations beyond the expected set: the corpus contains same-name decoys (another info(), other getAll()s), so extras expose reference implementations that match by name rather than by symbol. The corpus defines no namespace-level constant that code references — true of much modern PHP — so that kind awaits a define()-era corpus.