Fix KeyError in graph selector when using + operator with dbt-loom ex…

[award1230]

Summary

Fixes a KeyError when using the + (precursor) graph selector on a project that uses dbt-loom for cross-project references.

cc @pankajkoti @tatiana — This is a follow-up to your dbt-loom support in #2271. The external node skipping works great for basic rendering, but we hit a KeyError when combining it with the + graph selector. The + operator triggers select_node_precursors which traverses depends_on entries — and those can point to external nodes that were already filtered out during manifest loading. This code path wasn't exercised by the tests in #2271 since the example DAGs don't use graph selectors.

Problem

The dbt-loom support added in #2271 correctly skips external nodes (those without original_file_path) during manifest loading in load_from_dbt_manifest. However, local nodes still have depends_on entries pointing to these filtered-out external nodes.

When the + graph operator traverses upstream dependencies via select_node_precursors, it does nodes[node_id] on these external node IDs and raises a KeyError:

File "cosmos/dbt/selector.py", line 172, in select_node_precursors
    new_generation.update(set(nodes[node_id].depends_on))
                              ~~~~~^^^^^^^^^
KeyError: 'model.upstream_project.external_model'

Reproduction: Use select: ["+downstream_model"] in RenderConfig with LoadMode.DBT_MANIFEST on a project that uses dbt-loom with cross-project {{ ref('upstream_project', 'model_name') }} references.

Fix

Adds bounds checks in two locations in cosmos/dbt/selector.py:

1. GraphSelector.select_node_precursors (line 172): Skip node IDs not present in the nodes dict during upstream traversal
2. NodeSelector.select_nodes_ids_by_intersection (line 552): Skip external node IDs that were collected during graph traversal but don't exist in the nodes dict

This allows the + traversal to gracefully stop at project boundaries — the correct behavior for cross-project setups where external dependencies are managed by their own DAGs/task groups. This is consistent with how select_node_descendants already handles missing parents via defaultdict(set).

Test plan

• Added test_select_nodes_by_precursors_with_external_dependency — creates a graph where a local node's depends_on includes an external node ID not in the nodes dict, verifies + selector returns local nodes without KeyError
• All 166 existing selector tests pass
• All existing dbt-loom tests in test_graph.py pass

[tatiana]

@award1230 this is amazing, thank you very much for your contribution. It completely makes sense, and both the implementation & tests are spot on. I left inline minor comment - and I'm happy for us to merge this once you address. We could add this to the 1.13.1 release, planned for Monday

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 97.96%. Comparing base (9968648) to head (5e5d3ef).
⚠️ Report is 4 commits behind head on main.

Additional details and impacted files

@@           Coverage Diff           @@
##             main    #2389   +/-   ##
=======================================
  Coverage   97.96%   97.96%           
=======================================
  Files         102      102           
  Lines        6980     6983    +3     
=======================================
+ Hits         6838     6841    +3     
  Misses        142      142           

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[evanvolgas]

@tatiana I've found Github notifications to be a bit wonky sometimes. Cross posting here, in case:

#2394 is Alex's PR + the comments you asked for. Alex is out til next Monday but my team is hoping we can get his code merged as soon as possible.

Please let me know your thoughts!

[tatiana]

@evanvolgas thank you very much for following up on this. I really appreciate for the follow up PR with the comment. I'll go ahead and merge this PR, and if you can rebase your after, updating the PR description - it would be great! This way we acknowledge both of your contributions

[pankajkoti]

Nice, thanks a lot for testing out the addition to Cosmos and contribution of the fix @award1230 👏🏽