The prefetch recorded the rows its searches returned immediately, then went on to check for an empty result, stamp telemetry and format two markdown sections. Any failure in that tail is swallowed by the surrounding handler, which returns None and leaves `_prepare_query` building a prompt with no prefetch block at all -- so evidence could name conclusions the agent was never shown. Move the recording below the formatting, so what is recorded is what the returned context actually contains. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> |
||
|---|---|---|
| .. | ||
| __init__.py | ||
| test_evidence.py | ||
| test_model_config_usage.py | ||
| test_scope_preflight.py | ||
| test_structured_output.py | ||