Skip to content

Codebase intelligence for AI and humans: code health scores, auto-generated docs, git analytics, dead code detection, and architectural decisions via MCP.

fix(mcp): retain fallback_targets not named in emitted answer rows (#2995)

Rishabh Mishrac219425e7ad8
diff --git a/packages/server/src/repowise/server/mcp_server/tool_answer/projection.py b/packages/server/src/repowise/server/mcp_server/tool_answer/projection.py
index 88fabfe76..e6e0b99ff 100644
--- a/packages/server/src/repowise/server/mcp_server/tool_answer/projection.py
+++ b/packages/server/src/repowise/server/mcp_server/tool_answer/projection.py
@@ -392,7 +392,42 @@ def _slim_best_guesses(payload: dict[str, Any], facts: dict[str, Any]) -> bool:
return True
-def _default_shape(payload: dict[str, Any], question: str) -> None:
+def _prune_fallback_targets(
+ payload: dict[str, Any], raw_fallbacks: list[Any] | None
+) -> None:
+ """Retain fallback targets not already represented in emitted evidence rows."""
+ fallbacks = (
+ raw_fallbacks
+ if raw_fallbacks is not None
+ else payload.get("fallback_targets") or []
+ )
+ if not fallbacks:
+ payload.pop("fallback_targets", None)
+ return
+
+ emitted = {
+ path
+ for key in (
+ "citations",
+ "symbol_bodies",
+ "code_rationale",
+ "quotes",
+ "best_guesses",
+ "retrieval",
+ )
+ for row in payload.get(key) or []
+ if (path := _nav_path(row))
+ }
+ targets = [row for row in fallbacks if _nav_path(row) not in emitted]
+ if targets:
+ payload["fallback_targets"] = targets
+ else:
+ payload.pop("fallback_targets", None)
+
+
+def _default_shape(
+ payload: dict[str, Any], question: str, *, raw_fallbacks: list[Any] | None = None
+) -> None:
confidence = _shape_confidence(payload)
why = question.lstrip().lower().startswith("why")
if confidence == "high":
@@ -435,8 +470,8 @@ def _default_shape(payload: dict[str, Any], question: str) -> None:
else:
_keep(payload, "retrieval", 3)
payload.pop("candidates", None)
- if payload.get("best_guesses") or payload.get("retrieval") or payload.get("symbol_bodies"):
- payload.pop("fallback_targets", None)
+ _prune_fallback_targets(payload, raw_fallbacks)
+
if not str(payload.get("answer") or "").strip():
payload["answer"] = str(payload.get("note") or "No grounded answer was found.")
payload.setdefault(
@@ -497,7 +532,7 @@ def _serve_ranked_list(payload: dict[str, Any]) -> None:
{"path": _path(row), "lines": row.get("lines"), "comment": comment}
]
# Every cited path is a guess or a rationale row, so citations would repeat them.
- for key in ("citations", "note", "next_action_hint"):
+ for key in ("citations", "note", "next_action_hint", "fallback_targets"):
payload.pop(key, None)
if isinstance(payload.get("_meta"), dict):
payload["_meta"].pop("hint", None)
@@ -666,6 +701,7 @@ def _project(
) -> tuple[dict[str, Any], bool, bool]:
"""The projection, whether it slimmed ``best_guesses``, and whether it is the ranked list."""
payload = copy.deepcopy(raw)
+ raw_fallbacks = _unique(list(payload.get("fallback_targets") or []), lambda row: row)
totals = {
key: len(payload.get(key) or []) if isinstance(payload.get(key), list) else 0
for key in _COLLECTIONS
@@ -675,7 +711,7 @@ def _project(
facts = payload.pop("_candidate_file_facts", None)
slimmed = False
if not expanded:
- _default_shape(payload, question)
+ _default_shape(payload, question, raw_fallbacks=raw_fallbacks)
if _shape_confidence(payload) == "low":
slimmed = _slim_best_guesses(payload, facts if isinstance(facts, dict) else {})
bodies_cut = not expanded and _budget_keyless_bodies(payload, repo_root)
diff --git a/tests/unit/server/mcp/test_answer_projection.py b/tests/unit/server/mcp/test_answer_projection.py
index af3159508..850f8088c 100644
--- a/tests/unit/server/mcp/test_answer_projection.py
+++ b/tests/unit/server/mcp/test_answer_projection.py
@@ -733,3 +733,60 @@ def test_evidence_projection_only_gains_candidate_files():
assert after.pop("candidate_files") == _POOL[:5]
assert after == before
+
+
+def test_degraded_low_confidence_retains_trimmed_fallback_targets():
+ paths = [f"pkg/f{i}.py" for i in range(1, 6)]
+ raw = {
+ "answer": "...",
+ "citations": ["pkg/f1.py"],
+ "confidence": "low",
+ "retrieval_quality": "high",
+ "degraded": "synthesis-failed",
+ "fallback_targets": paths,
+ "retrieval": [{"path": p, "excerpt": "x" + p} for p in paths],
+ "code_rationale": [
+ {"path": "pkg/f1.py", "lines": [4, 5], "comment": "TCC rationale text"}
+ ],
+ "note": "model used its entire budget",
+ "_meta": {"degraded": "synthesis-failed"},
+ }
+ out = project_answer_payload(raw, question="where is LCOM4 computed")
+
+ assert [r["path"] for r in out["retrieval"]] == ["pkg/f1.py", "pkg/f2.py", "pkg/f3.py"]
+ assert out["fallback_targets"] == ["pkg/f4.py", "pkg/f5.py"]
+ assert out["fallback_targets_total"] == 5
+ assert out["fallback_targets_emitted"] == 2
+ assert out["retrieval_total"] == 5
+ assert out["retrieval_emitted"] == 3
+
+
+def test_high_confidence_drops_fallback_targets_with_overflow():
+ paths = [f"pkg/f{i}.py" for i in range(1, 6)]
+ raw = {
+ "answer": "Computed in pkg/f1.py.",
+ "citations": ["pkg/f1.py"],
+ "confidence": "high",
+ "retrieval_quality": "high",
+ "fallback_targets": paths,
+ "retrieval": [{"path": p, "excerpt": "x" + p} for p in paths],
+ }
+ out = project_answer_payload(raw, question="where is LCOM4 computed")
+ assert "fallback_targets" not in out
+ assert "retrieval" not in out
+
+
+def test_degraded_empty_retrieval_retains_full_fallback_targets():
+ paths = [f"pkg/f{i}.py" for i in range(1, 6)]
+ raw = {
+ "answer": "",
+ "citations": [],
+ "confidence": "low",
+ "degraded": "synthesis-failed",
+ "fallback_targets": paths,
+ "retrieval": [],
+ "note": "no local evidence matched",
+ }
+ out = project_answer_payload(raw, question="where is LCOM4 computed")
+ assert out["fallback_targets"] == paths
+ assert "retrieval" not in out