[nanami7777777]

Fixes #31

Three crashes / silent-wrong-result bugs found while running codegraph on a real-world project (326 modules). Two of them are hard crashes, one makes --file-path / --distance always report nothing.

---

1. AttributeError: 'NoneType' object has no attribute 'file' — module-level async def

create_objects_array()'s async branch assumed the token following async was always def, so async for / async with were parsed as function definitions (in async for x in ... the "name" becomes x). It also pushed None on the scope stack for a module-level async def instead of the AsyncFunction it had just created, so the next nested def / async token called _nest_function(None, ...) → ob.file → AttributeError.

Repro, three lines:

async def foo():
    return [x async for x in bar()]

Fix: validate the token after async (skip if it is not def), and assign cur_func before pushing it on the stack, mirroring the synchronous def branch.

The whole tree is parsed in one pass, so one such file made codegraph <dir> fail for the entire directory — 16 of 326 modules here.

2. TypeError: '<=' not supported between instances of 'int' and 'NoneType' — entities without an end line

Scopes popped by a sibling def / class (a one-line body such as def f(self) -> None: ... never emits INDENT/DEDENT) or still open at EOF never reach the DEDENT handler, so their endno stayed None. usage_graph() then compared entity[0] <= line <= entity[1] with entity[1] is None.

Fix: new _close_scopes() finalizes endno whenever a scope is popped; used in the async / def / class branches and once more after the token loop for anything left open at EOF.

On the project this took entities with endno is None from 74 → 0.

3. usage_graph() silently loses all import edges on a second call

get_imports_and_entities_lines() did code_objects[path].pop(-1) on the Import object, so a second usage_graph() call ran with an empty import set and therefore no cross-module edges. Since the CLI builds the graph (for the visualisation / CSV) before searching it, get_dependencies() always saw a degraded graph and --file-path / --distance printed nothing:

before:  {1: set(), 2: set()}
after:   {1: {llm.py, types.py, events.py, decorator.py, ...},
          2: {engine.py, context.py, records.py, gates.py, ...}}

Fix:

• keep the Import object on the list and skip it where entities are iterated, instead of popping it;
• get_dependencies() resolves a relative --file-path (graph keys are absolute, so codegraph src --file-path src/x.py never matched);
• an edge is followed past distance 1 when its bare "<module>.py" name maps to exactly one file. Ambiguous names are still reported but not traversed — there can be hundreds of __init__.py, and traversing them fans the BFS out to unrelated modules (distance 2 went from 8 meaningful files to ~120 without this guard).

---

Tests

tests/test_graph_generation.py:

• TestAsyncParsing — 6 cases (nested sync/async def, async for comprehension, async for statement, async with, module-level async def followed by a top-level def). 5 fail with the AttributeError before the fix.
• TestEntityEndLines — 4 cases (one-line body, last entity in module, scope closed by a sibling def, module-level async function).
• TestUsageGraphIdempotency — 2 cases (second usage_graph() equals the first; get_dependencies() still finds edges after usage_graph() was called).

46 tests pass; flake8 codegraph/ tests/ is clean.

Note on 1.3.0

#26 (Release/1.3.0) replaces the tokenizer parser with an AST-based one, which resolves 1 (and probably 2) for the default path — I verified the release/1.3.0 branch parses all 326 modules where 1.2.0 fails on 16. This PR targets the released 1.2.x line for users who cannot move yet; feel free to close it if 1.3.0 ships first. The legacy codegraph/parser.py on that branch still carries defects 1 and 2, so the same patch is worth applying there if that path stays reachable.

Known limitation left alone

Cross-file edges are derived from module.symbol usage plus basename matching, so a package import (from everos.component import embedding) collapses to a single __init__ node shared by every package. That is why distance-1 output contains a bare __init__.py. Fixing it properly needs real import resolution, which is a feature rather than a bug fix — out of scope here.