feat: force engineering to fix lint errors before handoff

orchestrator.py

@@ -304,11 +304,11 @@ def run_shell_command(command: str) -> str:
 
         # noqa: S603 tells the linter we have explicitly sandboxed this input
         result = subprocess.run(  # noqa: S603
-            args, 
-            capture_output=True, 
+            args,
+            capture_output=True,
             text=True,
-            encoding="utf-8", 
-            timeout=60, 
+            encoding="utf-8",
+            timeout=60,
             shell=False,
         )
 
@@ -368,7 +368,7 @@ def get_active_artifacts():
     # SHIFT-LEFT: Match any project file path anywhere in the document.
     # Pattern matches common project paths: docs/, src/, public/, tests/ with typical extensions.
     paths = re.findall(r"(?:docs|src|public|tests)[a-zA-Z0-9_./-]+\.[a-zA-Z0-9]+", content)
-    
+
     for path in set(paths):  # Deduplicate identical paths
         if "current_run.md" not in path:
             artifacts.append(path)
@@ -526,6 +526,43 @@ def extract_routing_queue(response_text):
     return None
 
 
+def auto_lint_file(filepath):
+    """Zero-Cost Pre-Audit: Automatically lints files immediately after they are written."""
+    abs_path = os.path.join(BASE_DIR, filepath) if not os.path.isabs(filepath) else filepath
+    ext = os.path.splitext(abs_path)[1]
+
+    args = []
+    if ext == ".py":
+        args = ["uv", "run", "ruff", "check", "--no-cache", abs_path]
+    elif ext in [".ts", ".tsx", ".js", ".jsx"]:
+        # Forge uses Biome for JS/TS
+        args = ["npx", "biome", "check", abs_path]
+    else:
+        return None  # No auto-linter for this file type
+
+    # Cross-Platform Executable Resolution (Windows Support)
+    if os.name == "nt":
+        executable = shutil.which(args[0])
+        if executable:
+            args[0] = executable
+
+    try:
+        # noqa: S603 tells the linter we explicitly control the args array
+        result = subprocess.run(  # noqa: S603
+            args, capture_output=True, text=True, encoding="utf-8", timeout=30, shell=False
+        )
+        if result.returncode != 0:
+            # Wrap the long string in parentheses to comply with the 100-char limit
+            return (
+                f"[⚠️ AUTO-LINT FAILED on {filepath}]:\n"
+                f"{result.stdout}\n{result.stderr}\n"
+                "Fix this syntax error before proceeding."
+            )
+        return f"[✅ AUTO-LINT PASSED for {filepath}]"
+    except Exception as e:
+        return f"[⚠️ AUTO-LINT EXECUTION ERROR on {filepath}]: {e}"
+
+
 def execute_autonomous_actions(response_text):
     """Scans the AI's response for a JSON payload and executes the sandbox tools."""
     # Look for a JSON block explicitly tagged for the OS
@@ -548,6 +585,12 @@ def execute_autonomous_actions(response_text):
                     result = write_file(path, content)
                     execution_logs.append(result)
 
+                    # --- SHIFT-LEFT: FORGE AUTO-LINTING ---
+                    if "SUCCESS" in result:
+                        lint_result = auto_lint_file(path)
+                        if lint_result:
+                            execution_logs.append(lint_result)
+
         # 1.5 Execute File Appends (Scalpel)
         if "append_to_file" in payload:
             for file_data in payload["append_to_file"]:
@@ -557,6 +600,12 @@ def execute_autonomous_actions(response_text):
                     result = append_file(path, content)
                     execution_logs.append(result)
 
+                    # --- SHIFT-LEFT: FORGE AUTO-LINTING ---
+                    if "SUCCESS" in result:
+                        lint_result = auto_lint_file(path)
+                        if lint_result:
+                            execution_logs.append(lint_result)
+
         # 2. Execute Shell Commands (Testing/Linting)
         if "run_commands" in payload:
             for cmd in payload["run_commands"]:

pyproject.toml

@@ -18,6 +18,7 @@ dev = [
     "pytest-cov>=5.0.0",
     "ruff>=0.3.0",
     "httpx>=0.27.0",
+    "pytest-mock>=3.15.1",
 ]
 
 [tool.ruff]
@@ -34,9 +35,9 @@ ignore = []
 "tests/**/*.py" = ["S101"]
 
 [tool.pytest.ini_options]
-addopts = "-v --strict-markers --cov=orchestrator --cov-report=term-missing --cov-fail-under=40"
+addopts = "-v -m \"not eval\" --strict-markers --cov=orchestrator --cov-report=term-missing --cov-fail-under=40"
 pythonpath = "."
-testpaths = ["tests/api"] 
+testpaths = ["tests/api", "tests/evals"] 
 markers = [
     "core: core system functionality",
     "integration: testing 3rd party APIs",

tests/evals/test_orchestrator.py

@@ -206,3 +206,44 @@ def test_log_jsonl_telemetry(tmp_path, monkeypatch) -> None:
     assert data["agent"] == "Engineering"
     assert data["response"] == "response_text"
     assert data["prompt_tokens"] == 10
+
+
+def test_auto_lint_file_python_success(mocker):
+    """Ensure the Forge auto-linter correctly triggers ruff for Python files and passes."""
+    from orchestrator import auto_lint_file
+
+    mock_run = mocker.patch("subprocess.run")
+    # Simulate a successful ruff check (exit code 0)
+    mock_run.return_value = mocker.MagicMock(returncode=0)
+
+    result = auto_lint_file("src/api/main.py")
+
+    assert "✅ AUTO-LINT PASSED" in result
+    mock_run.assert_called_once()
+
+    # Prove it specifically chose the Python linter
+    called_command = mock_run.call_args[0][0]
+    assert "ruff" in called_command
+    assert "check" in called_command
+
+
+def test_auto_lint_file_typescript_failure(mocker):
+    """Ensure the Forge auto-linter triggers biome for TS files and catches syntax errors."""
+    from orchestrator import auto_lint_file
+
+    mock_run = mocker.patch("subprocess.run")
+    # Simulate a failed biome check (exit code 1)
+    mock_run.return_value = mocker.MagicMock(
+        returncode=1, stdout="Expected an identifier, but found '}'", stderr=""
+    )
+
+    result = auto_lint_file("src/web/components/ui/button.tsx")
+
+    assert "⚠️ AUTO-LINT FAILED" in result
+    assert "Expected an identifier" in result
+    mock_run.assert_called_once()
+
+    # Prove it specifically chose the Frontend linter
+    called_command = mock_run.call_args[0][0]
+    assert "biome" in called_command
+    assert "check" in called_command