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(search): cap one file's line hits so test mocks cannot flood an identifier search (#3191)

Raghav Chamadiya987db713c89e
diff --git a/docs/agent/MCP_TOOLS.md b/docs/agent/MCP_TOOLS.md
index 5526c5c4b..8d33858e2 100644
--- a/docs/agent/MCP_TOOLS.md
+++ b/docs/agent/MCP_TOOLS.md
@@ -225,7 +225,7 @@ Hybrid search that routes by the shape of the query: identifiers search the symb
**Page rows** also carry `symbols`: up to three `name:line` entries (`Owner.member` for members) in that file whose names match the query's words. The first three page rows whose file has a matching line also carry `matched_lines`: a sample of up to three `{line, text}` source lines read from the live file that share the most query words, code lines before comments and imports. Unlike `lines`, it is never every match.
-An identifier or literal query also returns `lines` (`{path, line, kind, text}`, definitions first, at most 50): the `get_context` `references` of one to three exact symbols, else a live scan for the token (kind `definition` or `match`). `complete` is true only when they are every match (see `references` above); otherwise `reasons` says why.
+An identifier or literal query also returns `lines` (`{path, line, kind, text}`, definitions first, at most 50): the `get_context` `references` of one to three exact symbols, else a live scan for the token (kind `definition` or `match`). `complete` is true only when they are every match (see `references` above); otherwise `reasons` says why. Past 3 plain `reference` lines in one file, and past 50 lines, rows are counted in `lines_omitted_by_file` and recoverable from `_meta.omitted`; test files sort after production files.
```
search_codebase(query="GitIndexer index_repo")
diff --git a/packages/server/src/repowise/server/mcp_server/_line_hits.py b/packages/server/src/repowise/server/mcp_server/_line_hits.py
index acb7cc985..97c0bf4e5 100644
--- a/packages/server/src/repowise/server/mcp_server/_line_hits.py
+++ b/packages/server/src/repowise/server/mcp_server/_line_hits.py
@@ -18,6 +18,7 @@ import os
import re
import subprocess
import time
+from collections import Counter
from collections.abc import Container
from pathlib import Path
from typing import Any
@@ -28,10 +29,13 @@ from repowise.core.fs_walk import PRUNED_DIRS, walk_repo
from repowise.core.persistence.database import get_session
from repowise.core.persistence.models import GraphNode
from repowise.core.repo_config import load_repo_config
-from repowise.server.mcp_server._budget import register_post_enforce
+from repowise.core.test_paths import is_test_related_path
+from repowise.server.mcp_server._budget import OmissionCollector, register_post_enforce
from repowise.server.mcp_server._edit_sites import (
_MAX_FILE_BYTES,
+ MAX_REFERENCES_PER_FILE,
MAX_TEXT_CHARS,
+ _cap_references_per_file,
reference_edit_set,
)
from repowise.server.mcp_server._helpers import (
@@ -368,10 +372,31 @@ async def _attach(
rows, reasons = await asyncio.to_thread(_scan_listed, root, *literal)
else:
return
- rows.sort(key=lambda r: (_KIND_ORDER[r["kind"]], r["path"], r["line"]))
- if len(rows) > MAX_LINES and OVER_LINES not in reasons:
+ rows.sort(
+ key=lambda r: (
+ r["kind"] != "definition",
+ is_test_related_path(r["path"]),
+ _KIND_ORDER[r["kind"]],
+ r["path"],
+ r["line"],
+ )
+ )
+ kept = _cap_references_per_file(rows)
+ if len(kept) < len(rows):
+ reasons.append(
+ f"{len(rows) - len(kept)} references past {MAX_REFERENCES_PER_FILE} per file"
+ " not shown: lines_omitted_by_file; the full list is under _meta.omitted"
+ )
+ if len(kept) > MAX_LINES and OVER_LINES not in reasons:
reasons.append(OVER_LINES)
- response["lines"] = rows[:MAX_LINES]
+ response["lines"] = kept[:MAX_LINES]
+ if len(response["lines"]) < len(rows):
+ shown = {id(r) for r in response["lines"]}
+ hidden = [r for r in rows if id(r) not in shown]
+ collector = OmissionCollector("search_codebase", repo_root=ctx.path)
+ collector.add(f"{query} :: lines not shown", hidden)
+ collector.attach(response)
+ response["lines_omitted_by_file"] = dict(Counter(r["path"] for r in hidden))
response["complete"] = not reasons
if reasons:
response["reasons"] = reasons
diff --git a/tests/unit/server/mcp/test_search_line_hits.py b/tests/unit/server/mcp/test_search_line_hits.py
index e8f7bc1bc..5cf2fe780 100644
--- a/tests/unit/server/mcp/test_search_line_hits.py
+++ b/tests/unit/server/mcp/test_search_line_hits.py
@@ -303,3 +303,133 @@ def test_definition_patterns():
assert value.search(' "port": 47821,')
assert value.search("const PORT: number = 47821;")
assert value.search("connect(port=47821)") is None
+
+
+def _importer(session, rid: str, rel: str, target: str, names: str, language: str) -> None:
+ session.add_all(
+ [
+ GraphNode(
+ id=f"lh_file_{rel}",
+ repository_id=rid,
+ node_id=rel,
+ node_type="file",
+ language=language,
+ created_at=_NOW,
+ ),
+ GraphEdge(
+ id=f"lh_imp_{rel}",
+ repository_id=rid,
+ source_node_id=rel,
+ target_node_id=target,
+ edge_type="imports",
+ imported_names_json=names,
+ confidence=1.0,
+ created_at=_NOW,
+ ),
+ ]
+ )
+
+
+async def test_mock_heavy_python_test_collapses_into_recoverable_counts(tree, session, setup_mcp):
+ (tree / "tests").mkdir()
+ (tree / "tests" / "test_widget.py").write_text(
+ "from pkg.catalog import backend_of\n"
+ + "".join(f"mock.patch('pkg.catalog.backend_of', side_effect={i})\n" for i in range(20)),
+ encoding="utf-8",
+ )
+ _importer(session, setup_mcp, "tests/test_widget.py", "pkg/catalog.py", '["backend_of"]', "python")
+ await session.commit()
+
+ out = await search_codebase("backend_of", mode="symbol")
+
+ cap = _edit_sites.MAX_REFERENCES_PER_FILE
+ assert _rows(out) == [
+ ("pkg/catalog.py", 4, "definition"),
+ ("app/widget.py", 1, "import"),
+ ("app/widget.py", 6, "call"),
+ ("tests/test_widget.py", 1, "import"),
+ *(("tests/test_widget.py", n, "reference") for n in range(2, 2 + cap)),
+ ]
+ assert out["lines_omitted_by_file"] == {"tests/test_widget.py": 20 - cap}
+ assert out["complete"] is False
+ assert any("lines_omitted_by_file" in r for r in out["reasons"])
+ assert out["_meta"]["omitted"]["refs"]
+
+
+async def test_mock_heavy_typescript_test_sorts_after_the_real_user(tree, session, setup_mcp):
+ files = {
+ "web/a.ts": "export function pick(x: number): number { return x }\n",
+ "web/a.test.ts": "import { pick } from './a'\n" + "vi.mocked(pick).mockReturnValue(1)\n" * 30,
+ "web/b.ts": "import { pick } from './a'\nexport const run = () => [pick]\n",
+ }
+ for rel, text in files.items():
+ (tree / rel).parent.mkdir(parents=True, exist_ok=True)
+ (tree / rel).write_text(text, encoding="utf-8")
+ session.add(
+ GraphNode(
+ id="lh_ts_sym",
+ repository_id=setup_mcp,
+ node_id="web/a.ts::pick",
+ node_type="symbol",
+ name="pick",
+ file_path="web/a.ts",
+ kind="function",
+ language="typescript",
+ start_line=1,
+ end_line=1,
+ created_at=_NOW,
+ )
+ )
+ for rel in ("web/a.test.ts", "web/b.ts"):
+ _importer(session, setup_mcp, rel, "web/a.ts", '["pick"]', "typescript")
+ await session.commit()
+
+ out = await search_codebase("pick", mode="symbol")
+
+ paths = [p for p, _n, _k in _rows(out)]
+ assert ("web/b.ts", 2, "reference") in _rows(out)
+ assert paths.index("web/b.ts") < paths.index("web/a.test.ts")
+ assert paths.count("web/a.test.ts") == 1 + _edit_sites.MAX_REFERENCES_PER_FILE
+ assert out["lines_omitted_by_file"] == {
+ "web/a.test.ts": 30 - _edit_sites.MAX_REFERENCES_PER_FILE
+ }
+ assert out["complete"] is False
+
+
+async def test_call_sites_are_never_collapsed(tree, session, setup_mcp):
+ for i in range(4):
+ rel = f"svc/user{i}.py"
+ (tree / "svc").mkdir(exist_ok=True)
+ (tree / rel).write_text(
+ "from pkg.catalog import backend_of\n" + "backend_of(1)\n" * 6, encoding="utf-8"
+ )
+ _importer(session, setup_mcp, rel, "pkg/catalog.py", '["backend_of"]', "python")
+ session.add(
+ GraphEdge(
+ id=f"lh_call_{i}",
+ repository_id=setup_mcp,
+ source_node_id=f"{rel}::run",
+ target_node_id="pkg/catalog.py::backend_of",
+ edge_type="calls",
+ call_lines_json="[2, 3, 4, 5, 6, 7]",
+ confidence=0.95,
+ created_at=_NOW,
+ )
+ )
+ await session.commit()
+
+ out = await search_codebase("backend_of", mode="symbol")
+
+ calls = [r for r in _rows(out) if r[2] == "call" and r[0].startswith("svc/")]
+ assert len(calls) == 4 * 6
+ assert out["complete"] is True
+ assert "lines_omitted_by_file" not in out
+ assert "omitted" not in out["_meta"]
+
+
+def test_literal_rows_are_not_collapsed():
+ rows = [{"path": "a.py", "line": n, "kind": "match", "text": "x"} for n in range(5)]
+ assert _edit_sites._cap_references_per_file(rows) == rows