Compare commits
1 Commits
fix/781-js
...
fix/779
| Author | SHA1 | Date | |
|---|---|---|---|
|
|
c94d3cf3bf |
@@ -1,283 +0,0 @@
|
||||
"""
|
||||
JSON Repair for Tool-Call Pipeline — #781
|
||||
|
||||
Provides repair_and_load_json() that attempts to fix malformed JSON
|
||||
from LLM tool calls before falling back to full API retry.
|
||||
|
||||
Handles all common LLM JSON failure modes:
|
||||
- Truncation (missing closing braces/brackets)
|
||||
- Unquoted keys
|
||||
- Single quotes
|
||||
- Trailing commas
|
||||
- Markdown-wrapped JSON
|
||||
- Extra trailing text
|
||||
|
||||
Usage:
|
||||
from agent.json_repair import repair_and_load_json
|
||||
|
||||
args = repair_and_load_json(raw_args)
|
||||
if args is None:
|
||||
# Truly unrepairable — trigger retry
|
||||
pass
|
||||
"""
|
||||
|
||||
import json
|
||||
import re
|
||||
from typing import Any, Optional
|
||||
|
||||
try:
|
||||
from json_repair import repair_json
|
||||
HAS_JSON_REPAIR = True
|
||||
except ImportError:
|
||||
HAS_JSON_REPAIR = False
|
||||
|
||||
|
||||
def repair_and_load_json(text: str, default: Any = None) -> Any:
|
||||
"""
|
||||
Parse JSON, repairing common LLM malformations on failure.
|
||||
|
||||
Tries json.loads() first (fast path). On failure, attempts repair
|
||||
via json-repair library. Returns default if unrepairable.
|
||||
|
||||
Args:
|
||||
text: Raw JSON string (possibly malformed)
|
||||
default: Value to return if repair fails
|
||||
|
||||
Returns:
|
||||
Parsed JSON object, or default if unrepairable
|
||||
"""
|
||||
if not text or not isinstance(text, str):
|
||||
return default
|
||||
|
||||
text = text.strip()
|
||||
if not text:
|
||||
return default
|
||||
|
||||
# Fast path: try native parse first
|
||||
try:
|
||||
return json.loads(text)
|
||||
except (json.JSONDecodeError, TypeError, ValueError):
|
||||
pass
|
||||
|
||||
# Repair path
|
||||
repaired = _repair_json_string(text)
|
||||
if repaired is None:
|
||||
return default
|
||||
|
||||
try:
|
||||
return json.loads(repaired)
|
||||
except (json.JSONDecodeError, TypeError, ValueError):
|
||||
return default
|
||||
|
||||
|
||||
def _repair_json_string(text: str) -> Optional[str]:
|
||||
"""
|
||||
Attempt to repair a malformed JSON string.
|
||||
|
||||
Returns repaired string, or None if unrepairable.
|
||||
"""
|
||||
if not text:
|
||||
return None
|
||||
|
||||
# Strip markdown wrapping
|
||||
text = _strip_markdown(text)
|
||||
|
||||
# Try json-repair library if available
|
||||
if HAS_JSON_REPAIR:
|
||||
try:
|
||||
repaired = repair_json(text)
|
||||
if repaired and repaired != text:
|
||||
return repaired
|
||||
except Exception:
|
||||
pass
|
||||
|
||||
# Manual repair attempts
|
||||
repaired = _manual_repair(text)
|
||||
return repaired
|
||||
|
||||
|
||||
def _strip_markdown(text: str) -> str:
|
||||
"""Strip markdown code block wrapping."""
|
||||
# ```json ... ```
|
||||
match = re.match(r'^```(?:json)?\s*\n?(.*?)\n?```\s*$', text, re.DOTALL)
|
||||
if match:
|
||||
return match.group(1).strip()
|
||||
|
||||
# ` ... `
|
||||
match = re.match(r'^`(.+)`$', text, re.DOTALL)
|
||||
if match:
|
||||
return match.group(1).strip()
|
||||
|
||||
return text
|
||||
|
||||
|
||||
def _manual_repair(text: str) -> Optional[str]:
|
||||
"""Manual JSON repair heuristics."""
|
||||
original = text
|
||||
|
||||
# Remove trailing non-JSON text after valid JSON
|
||||
# Find the last } or ] that closes the root
|
||||
text = _extract_json(text)
|
||||
if text is None:
|
||||
return None
|
||||
|
||||
# Fix trailing commas
|
||||
text = re.sub(r',\s*([}\]])', r'\1', text)
|
||||
|
||||
# Fix single quotes -> double quotes (simple cases)
|
||||
# Only if there are no double quotes already
|
||||
if '"' not in text and "'" in text:
|
||||
text = text.replace("'", '"')
|
||||
|
||||
# Fix unquoted keys: {key: value} -> {"key": value}
|
||||
text = re.sub(r'([{,]\s*)([a-zA-Z_][a-zA-Z0-9_]*)\s*:', r'\1"\2":', text)
|
||||
|
||||
# Try to close unclosed braces/brackets
|
||||
text = _close_brackets(text)
|
||||
|
||||
return text if text != original else None
|
||||
|
||||
|
||||
def _extract_json(text: str) -> Optional[str]:
|
||||
"""Extract JSON from text that may have trailing content."""
|
||||
text = text.strip()
|
||||
|
||||
if not text:
|
||||
return None
|
||||
|
||||
# Find the start of JSON (first { or [)
|
||||
start_obj = text.find('{')
|
||||
start_arr = text.find('[')
|
||||
|
||||
if start_obj == -1 and start_arr == -1:
|
||||
return None
|
||||
|
||||
if start_arr == -1 or (start_obj != -1 and start_obj < start_arr):
|
||||
start = start_obj
|
||||
open_char, close_char = '{', '}'
|
||||
else:
|
||||
start = start_arr
|
||||
open_char, close_char = '[', ']'
|
||||
|
||||
# Find matching close, handling nesting
|
||||
depth = 0
|
||||
in_string = False
|
||||
escape_next = False
|
||||
|
||||
for i in range(start, len(text)):
|
||||
c = text[i]
|
||||
|
||||
if escape_next:
|
||||
escape_next = False
|
||||
continue
|
||||
|
||||
if c == '\\':
|
||||
escape_next = True
|
||||
continue
|
||||
|
||||
if c == '"' and not escape_next:
|
||||
in_string = not in_string
|
||||
continue
|
||||
|
||||
if in_string:
|
||||
continue
|
||||
|
||||
if c == open_char:
|
||||
depth += 1
|
||||
elif c == close_char:
|
||||
depth -= 1
|
||||
if depth == 0:
|
||||
return text[start:i+1]
|
||||
|
||||
# Unclosed — return from start to end, will be closed by _close_brackets
|
||||
return text[start:] if start > 0 else text
|
||||
|
||||
|
||||
def _close_brackets(text: str) -> str:
|
||||
"""Close unclosed braces and brackets."""
|
||||
open_count = 0
|
||||
close_count = 0
|
||||
in_string = False
|
||||
escape_next = False
|
||||
|
||||
for c in text:
|
||||
if escape_next:
|
||||
escape_next = False
|
||||
continue
|
||||
if c == '\\':
|
||||
escape_next = True
|
||||
continue
|
||||
if c == '"' and not escape_next:
|
||||
in_string = not in_string
|
||||
continue
|
||||
if in_string:
|
||||
continue
|
||||
if c == '{':
|
||||
open_count += 1
|
||||
elif c == '}':
|
||||
close_count += 1
|
||||
elif c == '[':
|
||||
open_count += 1
|
||||
elif c == ']':
|
||||
close_count += 1
|
||||
|
||||
# Close unclosed strings first
|
||||
if in_string:
|
||||
text += '"'
|
||||
|
||||
# Close unclosed braces/brackets
|
||||
diff = open_count - close_count
|
||||
if diff > 0:
|
||||
# Determine what to close based on last unclosed
|
||||
last_open_obj = text.rfind('{')
|
||||
last_open_arr = text.rfind('[')
|
||||
if last_open_obj > last_open_arr:
|
||||
text += '}' * diff
|
||||
else:
|
||||
text += ']' * diff
|
||||
|
||||
return text
|
||||
|
||||
|
||||
def repair_tool_call_arguments(arguments: str) -> tuple[str, bool]:
|
||||
"""
|
||||
Repair tool call arguments and return (repaired_args, was_repaired).
|
||||
|
||||
Specifically designed for the tool-call pipeline.
|
||||
"""
|
||||
if not arguments or not arguments.strip():
|
||||
return '{}', True
|
||||
|
||||
try:
|
||||
json.loads(arguments)
|
||||
return arguments, False # Already valid
|
||||
except (json.JSONDecodeError, TypeError, ValueError):
|
||||
pass
|
||||
|
||||
repaired = repair_and_load_json(arguments)
|
||||
if repaired is not None:
|
||||
return json.dumps(repaired), True
|
||||
|
||||
return '{}', True # Unrepairable — return empty dict
|
||||
|
||||
|
||||
# Metrics tracking
|
||||
_json_repairs_attempted = 0
|
||||
_json_repairs_succeeded = 0
|
||||
|
||||
|
||||
def get_repair_metrics() -> dict:
|
||||
"""Get JSON repair metrics."""
|
||||
return {
|
||||
"attempted": _json_repairs_attempted,
|
||||
"succeeded": _json_repairs_succeeded,
|
||||
"success_rate": _json_repairs_succeeded / _json_repairs_attempted if _json_repairs_attempted > 0 else 0,
|
||||
}
|
||||
|
||||
|
||||
def _track_repair(success: bool):
|
||||
"""Track repair attempt."""
|
||||
global _json_repairs_attempted, _json_repairs_succeeded
|
||||
_json_repairs_attempted += 1
|
||||
if success:
|
||||
_json_repairs_succeeded += 1
|
||||
20
tests/acp/conftest.py
Normal file
20
tests/acp/conftest.py
Normal file
@@ -0,0 +1,20 @@
|
||||
"""ACP test conftest — skip collection when acp extra not installed.
|
||||
|
||||
This conftest.py uses collect_ignore at module level to prevent
|
||||
pytest from trying to import test files that depend on acp.
|
||||
"""
|
||||
|
||||
try:
|
||||
import acp # noqa: F401
|
||||
except ImportError:
|
||||
# Tell pytest to skip this entire directory during collection
|
||||
collect_ignore = [
|
||||
"test_entry.py",
|
||||
"test_events.py",
|
||||
"test_mcp_e2e.py",
|
||||
"test_permissions.py",
|
||||
"test_server.py",
|
||||
"test_session.py",
|
||||
"test_tools.py",
|
||||
"test_auth.py",
|
||||
]
|
||||
@@ -15,6 +15,12 @@ PROJECT_ROOT = Path(__file__).parent.parent
|
||||
if str(PROJECT_ROOT) not in sys.path:
|
||||
sys.path.insert(0, str(PROJECT_ROOT))
|
||||
|
||||
# Register custom markers
|
||||
def pytest_configure(config):
|
||||
config.addinivalue_line("markers", "ssh: marks tests requiring SSH connectivity")
|
||||
config.addinivalue_line("markers", "integration: marks integration tests")
|
||||
config.addinivalue_line("markers", "slow: marks slow tests")
|
||||
|
||||
|
||||
@pytest.fixture(autouse=True)
|
||||
def _isolate_hermes_home(tmp_path, monkeypatch):
|
||||
@@ -119,3 +125,5 @@ def _enforce_test_timeout():
|
||||
yield
|
||||
signal.alarm(0)
|
||||
signal.signal(signal.SIGALRM, old)
|
||||
|
||||
|
||||
|
||||
@@ -1,136 +0,0 @@
|
||||
"""Tests for JSON repair module (#781)."""
|
||||
|
||||
import sys
|
||||
from pathlib import Path
|
||||
sys.path.insert(0, str(Path(__file__).parent.parent))
|
||||
|
||||
from agent.json_repair import repair_and_load_json, repair_tool_call_arguments, _strip_markdown, _extract_json
|
||||
|
||||
|
||||
class TestValidJSON:
|
||||
def test_valid_object(self):
|
||||
assert repair_and_load_json('{"key": "value"}') == {"key": "value"}
|
||||
|
||||
def test_valid_array(self):
|
||||
assert repair_and_load_json('[1, 2, 3]') == [1, 2, 3]
|
||||
|
||||
def test_empty_string_returns_default(self):
|
||||
assert repair_and_load_json('', default=42) == 42
|
||||
|
||||
def test_none_returns_default(self):
|
||||
assert repair_and_load_json(None, default=42) == 42
|
||||
|
||||
|
||||
class TestTrailingComma:
|
||||
def test_trailing_comma_object(self):
|
||||
assert repair_and_load_json('{"key": "value",}') == {"key": "value"}
|
||||
|
||||
def test_trailing_comma_array(self):
|
||||
assert repair_and_load_json('[1, 2, 3,]') == [1, 2, 3]
|
||||
|
||||
|
||||
class TestSingleQuotes:
|
||||
def test_single_quotes(self):
|
||||
result = repair_and_load_json("{'key': 'value'}")
|
||||
assert result == {"key": "value"}
|
||||
|
||||
|
||||
class TestUnquotedKeys:
|
||||
def test_unquoted_keys(self):
|
||||
result = repair_and_load_json('{key: "value"}')
|
||||
assert result == {"key": "value"}
|
||||
|
||||
|
||||
class TestTruncation:
|
||||
def test_missing_closing_brace(self):
|
||||
result = repair_and_load_json('{"name": "foo", "args": {"x": 1}')
|
||||
assert result is not None
|
||||
assert result["name"] == "foo"
|
||||
|
||||
def test_missing_closing_bracket(self):
|
||||
result = repair_and_load_json('[1, 2, 3')
|
||||
assert result == [1, 2, 3]
|
||||
|
||||
|
||||
class TestMarkdownWrapping:
|
||||
def test_json_codeblock(self):
|
||||
text = '```json\n{"key": "value"}\n```'
|
||||
assert repair_and_load_json(text) == {"key": "value"}
|
||||
|
||||
def test_plain_codeblock(self):
|
||||
text = '```\n{"key": "value"}\n```'
|
||||
assert repair_and_load_json(text) == {"key": "value"}
|
||||
|
||||
|
||||
class TestTrailingText:
|
||||
def test_extra_text_after_json(self):
|
||||
result = repair_and_load_json('{"result": true} some extra text')
|
||||
assert result == {"result": True}
|
||||
|
||||
|
||||
class TestStripMarkdown:
|
||||
def test_strip_json_block(self):
|
||||
assert _strip_markdown('```json\n{"a":1}\n```') == '{"a":1}'
|
||||
|
||||
def test_strip_plain_block(self):
|
||||
assert _strip_markdown('```\n{"a":1}\n```') == '{"a":1}'
|
||||
|
||||
|
||||
class TestExtractJSON:
|
||||
def test_extract_from_text(self):
|
||||
assert _extract_json('prefix {"key": 1} suffix') == '{"key": 1}'
|
||||
|
||||
def test_nested(self):
|
||||
assert _extract_json('{"a": {"b": 1}}') == '{"a": {"b": 1}}'
|
||||
|
||||
|
||||
class TestRepairToolCallArguments:
|
||||
def test_valid_args(self):
|
||||
args, repaired = repair_tool_call_arguments('{"x": 1}')
|
||||
assert args == '{"x": 1}'
|
||||
assert not repaired
|
||||
|
||||
def test_malformed_args(self):
|
||||
args, repaired = repair_tool_call_arguments("{x: 1}")
|
||||
assert repaired
|
||||
assert '"x"' in args
|
||||
|
||||
def test_empty_args(self):
|
||||
args, repaired = repair_tool_call_arguments('')
|
||||
assert args == '{}'
|
||||
assert repaired
|
||||
|
||||
|
||||
class TestFailureModes:
|
||||
def test_truncation_most_dangerous(self):
|
||||
# LLM cut off mid-string
|
||||
raw = '{"path": "test.py", "content": "hello'
|
||||
result = repair_and_load_json(raw)
|
||||
assert result is not None
|
||||
assert result.get("path") == "test.py"
|
||||
|
||||
def test_empty_arguments(self):
|
||||
# Auto-coerced to {}
|
||||
result = repair_and_load_json('')
|
||||
assert result is None # Returns default (None)
|
||||
|
||||
def test_non_json_text(self):
|
||||
result = repair_and_load_json('not json at all', default={})
|
||||
assert result == {}
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
import inspect
|
||||
tests = []
|
||||
for name, obj in inspect.getmembers(sys.modules[__name__]):
|
||||
if inspect.isclass(obj) and name.startswith('Test'):
|
||||
for method_name, method in inspect.getmembers(obj):
|
||||
if method_name.startswith('test_'):
|
||||
tests.append((f"{name}.{method_name}", method))
|
||||
|
||||
for name, test in tests:
|
||||
print(f"Running {name}...")
|
||||
test(None)
|
||||
print(" PASS")
|
||||
|
||||
print(f"\n{len(tests)} tests passed.")
|
||||
Reference in New Issue
Block a user