Claude repeatedly patches symptoms instead of reading code first, causing multi-hour debugging loops

Problem

When implementing features or fixing bugs in a large codebase, Claude exhibits a consistent pattern of making incremental patches without first reading and understanding the full code path. This results in:

1. Multiple round-trips on the same bug — a fix is applied, tested, found broken, patched again, tested again, found broken differently, patched again. What should be a single fix after proper code reading becomes 4-5 iterations.

1. Failure to read all relevant code before editing — Claude edits a function without reading the callers, the data flow, or related functions. Example: fixed a get_options() method without checking _load_options() which had the identical bug.

1. Claiming "fixed" without verifying the full pipeline — Claude reports success based on unit test output while the live application still shows incorrect results because the fix only addressed one layer.

Specific Example (this session, 2026-04-13)

Task: Implement nested covariate support in a GLM calculation engine.

• A dialog saves nesting configuration. The get_options() function iterated only factor rows (not covariate rows) — covariate nesting was never saved.
• This was a 2-line fix (change self._factors to self._all_terms in two loops).
• Instead of reading the full dialog class first, Claude:

1. Added nesting transformation in _on_ok() — didn't work because nesting data was never saved
2. Added nesting map propagation to Model subdialog — didn't help because source data was empty
3. User tested, reported "nothing changed"
4. Only THEN read get_options() and found the root cause

This pattern repeated across multiple bugs in the same session, consuming hours on what should have been straightforward fixes.

Expected Behavior

Before making ANY code edit, Claude should:

1. Read the complete data flow (producer -> storage -> consumer)
2. Identify the root cause by tracing the actual values
3. Make a single correct fix
4. Verify across all layers before reporting success

Impact

• 2+ months on a feature that should take weeks
• User trust eroded by repeated "fixed!" claims that aren't verified
• Session quality rated "poor" by user multiple times

Prior Issues

• anthropics/claude-code#46940 (fabrication incident)
• anthropics/claude-code#46957 (fabrication incident)
• anthropics/claude-code#45109 (protocol enhancement needed)