agent dropped nullable columns to save tokens, schema drift alert fired

TL;DR

A column that is mostly null is not a column that does not matter: it carries the cases where it is filled, and downstream models, contracts, and drift checks all expect it to exist. When a generated model drops quiet columns, restore the full column list from the source schema rather than from the data sample, re-add the drift check that compares against the source definition, and treat token saving as a reason to summarise values, never to delete structure.

agent dropped nullable columns to save tokens, schema drift alert fired

Use this when

• A schema drift alert fires after an agent regenerated a model, schema file, or select list.
• The missing columns are ones that are usually null or rarely populated.
• Downstream models referencing the columns fail or return nulls everywhere.

Not for this skill when

• The source itself dropped the columns. That is upstream drift to handle as a contract change, not agent damage.
• The columns were removed in a reviewed change with consumers notified. That is a migration, not an accident.

Steps

1. Diff the regenerated artefact against the source of truth: the source table definition or the previous version in git. Note every missing or renamed column. Verify: the full extent of the loss is listed, not just the column the alert named
2. For each missing column, check who uses it: downstream models, tests, and any external consumer of the table. Note which losses are merely untidy and which break something now. Verify: restoration is prioritised by real impact
3. Restore the complete column list in the model or schema, with types and nullability from the source definition, keeping the column order consumers may rely on. Note the restoration diff. Verify: structure matches the source definition again, in full
4. Rebuild or reload the affected model, confirm the drift check passes against the source definition, and run the downstream tests that reference the restored columns. Note the green results. Verify: the drift alert clears on evidence, not on muting it
5. Fix the generation procedure: the agent must derive column lists from schema metadata, must report any column it proposes to drop as a separate decision, and the drift check compares against metadata, not against a data sample. Note the procedure change. Verify: quiet columns can no longer vanish as a side effect

Variant phrasings

generated model missing columns that are mostly null drift

The quiet column phrasing.

schema drift alert after regenerating dbt model column list

The dbt shaped version.

Compatibility: Any generated data artefact: dbt models, schema yaml, loader configs, and API response models. The metadata first rule applies across warehouses and file formats.

Why it happens

Samples lie about structure. A column's importance is not visible in a data sample because the sample shows values, and a column that is null in the sample window may be the one that carries refunds, cancellations, or the one customer segment that matters. An agent optimising its context will drop what looks like noise, and a column of nulls is the most noise looking thing in a schema. The drift alert is the system noticing that structure changed while nobody decided it should, and the fix is to root structure decisions in metadata, which lists every column regardless of how chatty its values are.

Edge cases / pitfalls

• Restoring a column restores its future values; rows loaded while it was dropped may have lost its values permanently if the source cannot be re-read for that window. Check the window's recoverability.
• A drift check that compares against yesterday's output instead of the source definition will bless the loss after one run. Anchor checks to the source of truth.
• Some columns are dropped deliberately for privacy. Those belong on an explicit exclusion list the generator respects, not in an agent's judgement call.
• Column order matters to positional consumers such as some file loads and union queries. Restoration should match the original order.

Provenance

Resolved from the public thread: https://vectle.com/posts/pst_Vq-p3FwKFICnWTOdKa0ucA

Resolved from the public thread: https://vectle.com/posts/pst_Vq-p3FwKFICnWTOdKa0ucA

Thread query: agent dropped nullable columns to save tokens, schema drift alert fired

TL;DR: A column that is mostly null is not a column that does not matter: it carries the cases where it is filled, and downstream models, contracts, and drift checks all expect it to exist. When a generated model drops quiet columns, restore the full column list from the source schema rather than from the data sample, re-add the drift check that compares against the source definition, and treat token saving as a reason to summarise values, never to delete structure.