📣 New Release: codeanalyzer-python v0.3.0

[rahlk]

Changelog

Added

• --analysis-level {1,2} (reintroduced): 1 is symbol table + Jedi call graph, 2 adds the PyCG call graph.
• Coupling-aware PyCG sharding (--pycg-shard) so level 2 scales to large apps. Shards are chosen by Jedi module coupling (strongly-connected-component condensation, so import cycles never split, plus Louvain community detection) instead of a flat file count, so few call edges are severed between shards. PyCG runs on each shard inside a symlink mini-project that bounds it to that shard's files. Ray-parallel.
• Iterative decomposition of runaway shards. PyCG's fixpoint can diverge on heavy metaclass / mixin code. A shard that hits the wall-clock timeout is re-sharded at half the budget and re-run, down to a floor (10 files). The residue that still diverges, or an atomic cycle that will not split, falls back to Jedi-only coverage. On the Odoo benchmark (1028 modules) this recovered 22210 PyCG edges versus 17149 for the best uniform ceiling, losing only 20 of 1028 files instead of a whole shard.
• New flags: --pycg-shard-strategy {jedi,package} (default jedi), --pycg-shard-ceiling (default 100, the starting per-shard budget), --pycg-shard-timeout (default 120, per-shard wall clock), --pycg-max-iter (default 50, caps PyCG fixpoint passes so a divergent shard returns a partial graph instead of hanging).

Changed

• BREAKING: CodeQL is replaced by PyCG as the level 2 call graph backend. --codeql/--no-codeql is removed in favor of --analysis-level. Edge provenance literal codeql becomes pycg. New dependency: pycg (Apache 2.0). The codeanalyzer.semantic_analysis.codeql package is removed.

Fixed

• The shard planner keys its module graph by file path. PyModule.module_name is only the file stem (every __init__.py, models.py, ...), so keying by name collided and silently dropped files from shards.
• PyCG no longer follows imports into an in-tree dependency venv (e.g. .codeanalyzer/) during the whole-project level 2 run: it runs in a symlink mini-project whose root holds project source only, so dependencies resolve outside the bound and stay ghost nodes.
• _uv_bin uses only the vendored uv from the uv PyPI dependency and no longer falls back to a uv on PATH, so the analyzer always uses the pinned binary. Returns None only if the package is missing, in which case callers fall back to pip.

---

Install codeanalyzer-python v0.3.0

Shell script (installs the canpy CLI via uv / pipx / pip):

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

PyPI:

pip install codeanalyzer-python==0.3.0

For the optional live Neo4j push (--emit neo4j --neo4j-uri ...):

pip install 'codeanalyzer-python[neo4j]==0.3.0'

Download

File	Description
codeanalyzer_python-0.3.0-py3-none-any.whl	Python wheel
codeanalyzer_python-0.3.0.tar.gz	Source distribution
canpy-installer.sh	Shell installer (uv / pipx / pip)
schema.json	Neo4j schema contract

What's Changed

🛠 Other Changes

• PyCG sharding: Jedi planner + adaptive decomposition of runaways by @rahlk in PyCG sharding: Jedi planner + adaptive decomposition of runaways #49
• Level 2: replace CodeQL with PyCG, add coupling-aware adaptive sharding by @rahlk in Level 2: replace CodeQL with PyCG, add coupling-aware adaptive sharding #50
• Merge pull request PyCG sharding: Jedi planner + adaptive decomposition of runaways #49 from codellm-devkit/feat/jedi-shard-planner by @rahlk in Merge pull request #49 from codellm-devkit/feat/jedi-shard-planner #51
• fix(venv): use only the vendored uv, drop system-PATH fallback by @rahlk in fix(venv): use only the vendored uv, drop system-PATH fallback #52

Full Changelog: v0.2.1...v0.3.0