From 15cdf01f58c7bc1c06ae574d167c558a87d94bf0 Mon Sep 17 00:00:00 2001
From: Eric Rakestraw <eric@rakestraw.us>
Date: Wed, 22 Apr 2026 11:36:24 -0500
Subject: [PATCH] Bugfixes in the LLM retry loop

---
 src/audita/pipeline.py |   8 ++-
 tests/test_pipeline.py | 113 ++++++++++++++++++++++++++++++++++++++++-
 2 files changed, 115 insertions(+), 6 deletions(-)

diff --git a/src/audita/pipeline.py b/src/audita/pipeline.py
index ed44189..5235f2e 100644
--- a/src/audita/pipeline.py
+++ b/src/audita/pipeline.py
@@ -195,15 +195,13 @@ def _run_correction_stage(
         )
         working = application_result.transcript
 
-        for correction_id in application_result.applied_ids:
-            unresolved_retry_skips.pop(correction_id, None)
-        for correction_id in application_result.ignored_ids:
-            unresolved_retry_skips.pop(correction_id, None)
+        next_retry_skips: Dict[int, SkippedCorrection] = {}
         for skipped in application_result.skipped:
             if _is_retryable_skip(skipped, working):
-                unresolved_retry_skips[skipped.id] = skipped
+                next_retry_skips[skipped.id] = skipped
             else:
                 final_nonretry_skips.append(skipped)
+        unresolved_retry_skips = next_retry_skips
 
         pass_summaries.append(
             {
diff --git a/tests/test_pipeline.py b/tests/test_pipeline.py
index 88f7889..b4b8fae 100644
--- a/tests/test_pipeline.py
+++ b/tests/test_pipeline.py
@@ -310,6 +310,110 @@ def test_pipeline_retries_skipped_segment_and_cleans_work_dir_when_fixed(tmp_pat
     assert list((tmp_path / "work").iterdir()) == []
 
 
+def test_empty_glossary_retry_response_stops_retrying_segment(tmp_path):
+    first_pass = CorrectionCandidate(
+        id=1,
+        original_text="Contia",
+        corrected_text="Chauntea",
+        confidence=0.95,
+    )
+    fake_client = FakeLLMClient(
+        [
+            CorrectionSet(corrections=[first_pass]),
+            CorrectionSet(corrections=[]),
+            CorrectionSet(corrections=[]),
+            CorrectionSet(corrections=[]),
+        ]
+    )
+
+    revised = process_transcript(
+        _transcript(),
+        _glossary(),
+        _config(tmp_path, glossary_max_llm_passes=3),
+        llm_client=fake_client,
+    )
+
+    assert fake_client.calls == 3
+    assert "readability corrections" in fake_client.messages[2][1]["content"]
+    assert revised[0].text == "I ask Chontia."
+    assert list((tmp_path / "work").iterdir()) == []
+
+
+def test_empty_grammar_retry_response_stops_retrying_segment(tmp_path):
+    transcript = parse_source_transcript_json(
+        """
+        [
+          {"speaker": "Eric", "start": 0.0, "end": 1.0, "text": "there and there."}
+        ]
+        """
+    )
+    repeated_span = CorrectionCandidate(
+        id=1,
+        original_text="there",
+        corrected_text="their",
+        confidence=0.95,
+    )
+    fake_client = FakeLLMClient(
+        [
+            CorrectionSet(corrections=[]),
+            CorrectionSet(corrections=[repeated_span]),
+            CorrectionSet(corrections=[]),
+            CorrectionSet(corrections=[]),
+        ]
+    )
+
+    revised = process_transcript(
+        transcript,
+        _glossary(),
+        _config(tmp_path, grammar_max_llm_passes=3),
+        llm_client=fake_client,
+    )
+
+    assert fake_client.calls == 3
+    assert revised[0].text == "there and there."
+    assert list((tmp_path / "work").iterdir()) == []
+
+
+def test_retry_pass_with_new_skip_schedules_following_pass(tmp_path):
+    first_pass = CorrectionCandidate(
+        id=1,
+        original_text="Contia",
+        corrected_text="Chauntea",
+        confidence=0.95,
+    )
+    second_pass = CorrectionCandidate(
+        id=1,
+        original_text="Chantia",
+        corrected_text="Chauntea",
+        confidence=0.95,
+    )
+    third_pass = CorrectionCandidate(
+        id=1,
+        original_text="Chontia",
+        corrected_text="Chauntea",
+        confidence=0.95,
+    )
+    fake_client = FakeLLMClient(
+        [
+            CorrectionSet(corrections=[first_pass]),
+            CorrectionSet(corrections=[second_pass]),
+            CorrectionSet(corrections=[third_pass]),
+            CorrectionSet(corrections=[]),
+        ]
+    )
+
+    revised = process_transcript(
+        _transcript(),
+        _glossary(),
+        _config(tmp_path, glossary_max_llm_passes=3),
+        llm_client=fake_client,
+    )
+
+    assert fake_client.calls == 4
+    assert revised[0].text == "I ask Chauntea."
+    assert list((tmp_path / "work").iterdir()) == []
+
+
 def test_pipeline_retry_prompt_contains_only_valid_deduped_ids(tmp_path):
     first_bad = CorrectionCandidate(
         id=1,
@@ -358,10 +462,16 @@ def test_pipeline_writes_stage_metadata_for_unresolved_retries(tmp_path):
         corrected_text="Chauntea",
         confidence=0.95,
     )
+    second_pass = CorrectionCandidate(
+        id=1,
+        original_text="Chantia",
+        corrected_text="Chauntea",
+        confidence=0.95,
+    )
     fake_client = FakeLLMClient(
         [
             CorrectionSet(corrections=[first_pass]),
-            CorrectionSet(corrections=[]),
+            CorrectionSet(corrections=[second_pass]),
             CorrectionSet(corrections=[]),
         ]
     )
@@ -392,6 +502,7 @@ def test_pipeline_writes_stage_metadata_for_unresolved_retries(tmp_path):
     assert [item["pass_number"] for item in metadata["stages"][0]["passes"]] == [1, 2]
     assert metadata["stages"][0]["passes"][0]["retry_segment_count"] == 1
     assert metadata["stages"][0]["passes"][1]["retry_pass"] is True
+    assert metadata["stages"][0]["passes"][1]["retry_segment_count"] == 1
 
 
 def test_grammar_stage_runs_after_glossary_and_sees_corrected_text(tmp_path):