codeanalyzer-typescript v1.5.0

[rahlk]

codeanalyzer-typescript v1.5.0

Four correctness fixes, two of which change values you may have stored. Neo4j contract unchanged at 2.0.0. Read the first section before upgrading a pipeline that keeps span.bytes or can:// ids of merged declarations.

⚠️ Value changes (minor bump, but read this)

span.bytes are now UTF-8 byte offsets (#189, closes #179)

They were UTF-16 char offsets — ts-morph's positions — while the schema, the canonical keystone and codeanalyzer-python all mean bytes. The Neo4j projection already sliced a Buffer by them, so :TSCallable.code ran short by exactly the multibyte surplus inside the span (one em dash → the closing \n} missing). Now every node at every level — declarations, call/config_access body nodes, L3 statements and @entry/@exit, the module span, yaml ConfigKey spans — carries real byte offsets, and Buffer.from(source).subarray(lo, hi) reproduces its text, the same rule as python's source.encode()[lo:hi].

Migration: slice with bytes, not chars. Files with no non-ASCII character are byte-identical before and after; on any other file every span after the first multibyte character moves. A stored span.bytes from ≤ 1.4.0 must be recomputed.

Declaration merging mints one id per facet (#190, closes #177)

A value and a type of one name (const TableOption = () => … + interface TableOption) shared one can:// id, so the graph MERGEd them onto a node with two labels and the wrong kind; a type/type merge (class C + interface C) dropped the class from analysis.json entirely. Now the value facet keeps the bare id and the colliding type becomes …/TableOption#type; in a type/type merge the first facet keeps the bare key and id and each later one is keyed Name#<kind> (C#interface) with id …/C#interface. Signatures never move; extends/implements resolve to the type facet.

Migration: only ids that collided before change; a project without declaration merging is byte-identical (verified against main). If you stored the id of a merged type, expect the suffix.

Fixes

Large repositories no longer die in finalizeAnalysis (#188, closes #180) — the wire copy was one structuredClone of the whole envelope, which crosses Bun's serialization ceiling on a 13k-file repository (a TypeError on 1.3.0; measured on Bun 1.3.14: a ~2 GiB object aborts the process with no exception). The clone is per module now — megabytes, never gigabytes — and the rest of the root is shared by reference.

A missing input path is an error (#187, closes #181) — -i /does/not/exist used to write a schema-valid analysis.json for an empty application and exit 0. Now: one line on stderr naming the path, exit 1, nothing written; --list-programs refuses the same way. An empty but real directory still exits 0 with an empty symbol_table — that is a real answer. The reported "FATAL then exit 0" could not be reproduced on main (the FATAL path exits 1); #188 removes its only known trigger.

Also: export { type X } per-specifier type-only flags are honoured; heritage of a merged symbol resolves to its type declaration rather than its first declaration.

Upgrade

pip install -U "codeanalyzer-typescript==1.5.0"

brew upgrade codellm-devkit/homebrew-tap/codeanalyzer-typescript

curl --proto '=https' --tlsv1.2 -LsSf https://github.com/codellm-devkit/codeanalyzer-typescript/releases/latest/download/cants-installer.sh | sh

A warm .codeanalyzer cache from 1.4.0 is invalidated automatically.

Links

• Release v1.5.0 — prebuilt binaries, wheels and the Neo4j schema contract
• Full diff, v1.4.0...v1.5.0
• Design records: span.bytes are bytes, declaration merging ids