| 1 | from __future__ import annotations |
| 2 | |
| 3 | from pathlib import Path |
| 4 | import subprocess |
| 5 | import sys |
| 6 | from types import SimpleNamespace |
| 7 | |
| 8 | import pytest |
| 9 | |
| 10 | from extensions.python.system_prompt import _11_tools_prompt, _13_skills_prompt |
| 11 | from helpers import files, mcp_handler, responses_tools, tool_policy |
| 12 | from helpers.errors import RepairableException |
| 13 | from plugins._tool_access.extensions.python.tool_execute_before._10_enforce_tool_policy import ( |
| 14 | EnforceToolPolicy, |
| 15 | ) |
| 16 | |
| 17 | |
| 18 | class _Context: |
| 19 | def get_data(self, key: str, recursive: bool = True): |
| 20 | return None |
| 21 | |
| 22 | |
| 23 | class _Agent: |
| 24 | def __init__(self, prompt_root: Path, profile: str = "researcher") -> None: |
| 25 | self.prompt_root = prompt_root |
| 26 | self.config = SimpleNamespace(profile=profile) |
| 27 | self.context = _Context() |
| 28 | self.data: dict = {} |
| 29 | |
| 30 | def read_prompt(self, basename: str, **kwargs) -> str: |
| 31 | content = (self.prompt_root / basename).read_text(encoding="utf-8") |
| 32 | for key, value in kwargs.items(): |
| 33 | content = content.replace("{{" + key + "}}", str(value)) |
| 34 | return content |
| 35 | |
| 36 | def get_data(self, key: str): |
| 37 | return self.data.get(key) |
| 38 | |
| 39 | |
| 40 | class _NoMCPTools: |
| 41 | def get_tools(self): |
| 42 | return [] |
| 43 | |
| 44 | |
| 45 | def _write_prompt(root: Path, basename: str, content: str) -> None: |
| 46 | (root / basename).write_text(content.strip() + "\n", encoding="utf-8") |
| 47 | |
| 48 | |
| 49 | def _prompt_paths(root: Path): |
| 50 | def get_paths(agent, *parts, **kwargs): |
| 51 | return [str(root)] if parts and parts[0] == "prompts" else [] |
| 52 | |
| 53 | return get_paths |
| 54 | |
| 55 | |
| 56 | def _custom_policy(*, default: str, mcp_default: str = "allow", allowed=(), blocked=()): |
| 57 | return { |
| 58 | "mode": "custom", |
| 59 | "default": default, |
| 60 | "mcp_default": mcp_default, |
| 61 | "allowed": list(allowed), |
| 62 | "blocked": list(blocked), |
| 63 | } |
| 64 | |
| 65 | |
| 66 | def test_agent_import_does_not_cycle_through_tool_policy() -> None: |
| 67 | subprocess.run( |
| 68 | [sys.executable, "-c", "import agent"], |
| 69 | cwd=Path(__file__).parents[1], |
| 70 | check=True, |
| 71 | ) |
| 72 | |
| 73 | |
| 74 | @pytest.fixture |
| 75 | def local_prompt_agent(monkeypatch, tmp_path: Path) -> _Agent: |
| 76 | _write_prompt(tmp_path, "agent.system.tools.md", "TOOLS\n{{tools}}") |
| 77 | _write_prompt( |
| 78 | tmp_path, |
| 79 | "agent.system.tool.allowed.md", |
| 80 | """### allowed |
| 81 | Allowed description |
| 82 | Keyboard input remains documented. |
| 83 | Do not call the `blocked` tool from here. |
| 84 | {"tool_name":"allowed","tool_args":{}}""", |
| 85 | ) |
| 86 | _write_prompt( |
| 87 | tmp_path, |
| 88 | "agent.system.tool.blocked.md", |
| 89 | '### blocked\nBlocked description\n{"tool_name":"blocked","tool_args":{}}', |
| 90 | ) |
| 91 | monkeypatch.setattr(tool_policy.subagents, "get_paths", _prompt_paths(tmp_path)) |
| 92 | monkeypatch.setattr( |
| 93 | "plugins._model_config.helpers.model_config.get_chat_model_config", |
| 94 | lambda agent: {"vision": False}, |
| 95 | ) |
| 96 | monkeypatch.setattr(responses_tools, "_mcp_tools", lambda agent: []) |
| 97 | return _Agent(tmp_path) |
| 98 | |
| 99 | |
| 100 | @pytest.mark.asyncio |
| 101 | async def test_text_tool_prompt_omits_blocked_tool_and_description( |
| 102 | monkeypatch, local_prompt_agent: _Agent |
| 103 | ) -> None: |
| 104 | monkeypatch.setattr( |
| 105 | tool_policy, |
| 106 | "get_policy", |
| 107 | lambda agent: _custom_policy(default="allow", blocked=["local:blocked"]), |
| 108 | ) |
| 109 | |
| 110 | prompt = await _11_tools_prompt.build_prompt(local_prompt_agent) |
| 111 | |
| 112 | assert "Allowed description" in prompt |
| 113 | assert "Keyboard input remains documented." in prompt |
| 114 | assert "Do not call" not in prompt |
| 115 | assert "blocked" not in prompt.lower() |
| 116 | assert "Blocked description" not in prompt |
| 117 | |
| 118 | |
| 119 | def test_provider_native_schemas_omit_blocked_local_tool( |
| 120 | monkeypatch, local_prompt_agent: _Agent |
| 121 | ) -> None: |
| 122 | monkeypatch.setattr( |
| 123 | tool_policy, |
| 124 | "get_policy", |
| 125 | lambda agent: _custom_policy(default="allow", blocked=["local:blocked"]), |
| 126 | ) |
| 127 | |
| 128 | tools, _name_map = responses_tools.build_responses_function_tools( |
| 129 | local_prompt_agent |
| 130 | ) |
| 131 | |
| 132 | assert [tool["name"] for tool in tools] == ["allowed"] |
| 133 | |
| 134 | |
| 135 | def test_inherited_prompt_filter_skips_tool_inventory(monkeypatch, tmp_path: Path): |
| 136 | agent = _Agent(tmp_path) |
| 137 | prompt = "### shell\nRun a command." |
| 138 | policy_reads = 0 |
| 139 | |
| 140 | def inherited_policy(_agent): |
| 141 | nonlocal policy_reads |
| 142 | policy_reads += 1 |
| 143 | return { |
| 144 | "mode": "inherit", |
| 145 | "default": "allow", |
| 146 | "mcp_default": "allow", |
| 147 | "allowed": [], |
| 148 | "blocked": [], |
| 149 | } |
| 150 | |
| 151 | monkeypatch.setattr(tool_policy, "get_policy", inherited_policy) |
| 152 | monkeypatch.setattr( |
| 153 | tool_policy, |
| 154 | "_policy_tool_names", |
| 155 | lambda _agent: pytest.fail("inherited policy inventoried tools"), |
| 156 | ) |
| 157 | |
| 158 | assert tool_policy.filter_tool_prompt( |
| 159 | agent, "agent.system.tool.shell.md", prompt |
| 160 | ) == prompt |
| 161 | assert policy_reads == 1 |
| 162 | |
| 163 | |
| 164 | def test_required_response_survives_default_block(monkeypatch, tmp_path: Path) -> None: |
| 165 | _write_prompt( |
| 166 | tmp_path, |
| 167 | "agent.system.tool.response.md", |
| 168 | '### response\nfinal answer\n{"tool_name":"response","tool_args":{"text":"done"}}', |
| 169 | ) |
| 170 | monkeypatch.setattr(tool_policy.subagents, "get_paths", _prompt_paths(tmp_path)) |
| 171 | monkeypatch.setattr( |
| 172 | tool_policy, |
| 173 | "get_policy", |
| 174 | lambda agent: _custom_policy(default="block", blocked=["local:response"]), |
| 175 | ) |
| 176 | monkeypatch.setattr( |
| 177 | mcp_handler.MCPConfig, |
| 178 | "get_for_agent", |
| 179 | lambda agent: _NoMCPTools(), |
| 180 | ) |
| 181 | agent = _Agent(tmp_path) |
| 182 | |
| 183 | decision = tool_policy.resolve_tool(agent, "response") |
| 184 | |
| 185 | assert decision.allowed is True |
| 186 | assert decision.source == "framework-required" |
| 187 | assert tool_policy.get_tool_catalog(agent) == [] |
| 188 | |
| 189 | |
| 190 | def test_tool_and_mcp_defaults_are_independent(monkeypatch, tmp_path: Path) -> None: |
| 191 | agent = _Agent(tmp_path) |
| 192 | monkeypatch.setattr( |
| 193 | tool_policy, |
| 194 | "get_policy", |
| 195 | lambda _agent: _custom_policy( |
| 196 | default="block", |
| 197 | mcp_default="allow", |
| 198 | allowed=["local:pinned"], |
| 199 | blocked=["mcp:docs:delete"], |
| 200 | ), |
| 201 | ) |
| 202 | |
| 203 | assert tool_policy.resolve_tool(agent, "shell", canonical_id="local:shell").allowed is False |
| 204 | assert tool_policy.resolve_tool(agent, "read", canonical_id="mcp:docs:read").allowed is True |
| 205 | assert tool_policy.resolve_tool(agent, "pinned", canonical_id="local:pinned").allowed is True |
| 206 | assert tool_policy.resolve_tool(agent, "delete", canonical_id="mcp:docs:delete").allowed is False |
| 207 | |
| 208 | |
| 209 | def test_catalog_comes_from_executable_tools_not_prompt_names( |
| 210 | monkeypatch, tmp_path: Path |
| 211 | ) -> None: |
| 212 | prompt_root = tmp_path / "prompts" |
| 213 | tool_root = tmp_path / "tools" |
| 214 | prompt_root.mkdir() |
| 215 | tool_root.mkdir() |
| 216 | _write_prompt( |
| 217 | prompt_root, |
| 218 | "agent.system.tool.actual.md", |
| 219 | "### actual\nActual description", |
| 220 | ) |
| 221 | _write_prompt( |
| 222 | prompt_root, |
| 223 | "agent.system.tool.prompt_only.md", |
| 224 | "### prompt_only\nNo executable implementation", |
| 225 | ) |
| 226 | (tool_root / "actual.py").write_text("class Actual: pass\n", encoding="utf-8") |
| 227 | (tool_root / "response.py").write_text("class Response: pass\n", encoding="utf-8") |
| 228 | |
| 229 | def get_paths(agent, *parts, **kwargs): |
| 230 | if parts[0] == "prompts": |
| 231 | return [str(prompt_root)] |
| 232 | if len(parts) == 1: |
| 233 | return [str(tool_root)] |
| 234 | return [str(tool_root / parts[1])] |
| 235 | |
| 236 | monkeypatch.setattr(tool_policy.subagents, "get_paths", get_paths) |
| 237 | monkeypatch.setattr( |
| 238 | mcp_handler.MCPConfig, |
| 239 | "get_for_agent", |
| 240 | lambda agent: _NoMCPTools(), |
| 241 | ) |
| 242 | monkeypatch.setattr( |
| 243 | tool_policy, |
| 244 | "get_policy", |
| 245 | lambda agent: { |
| 246 | "mode": "inherit", |
| 247 | "default": "allow", |
| 248 | "allowed": [], |
| 249 | "blocked": [], |
| 250 | }, |
| 251 | ) |
| 252 | |
| 253 | catalog = tool_policy.get_tool_catalog(_Agent(prompt_root)) |
| 254 | |
| 255 | assert [item["id"] for item in catalog] == ["local:actual"] |
| 256 | assert catalog[0]["description"] == "Actual description" |
| 257 | |
| 258 | |
| 259 | def test_catalog_keeps_installed_remote_tools_without_live_connector( |
| 260 | monkeypatch, tmp_path: Path |
| 261 | ) -> None: |
| 262 | tool_root = tmp_path / "tools" |
| 263 | tool_root.mkdir() |
| 264 | (tool_root / "code_execution_remote.py").write_text("", encoding="utf-8") |
| 265 | (tool_root / "shell.py").write_text("", encoding="utf-8") |
| 266 | monkeypatch.setattr( |
| 267 | tool_policy.subagents, |
| 268 | "get_paths", |
| 269 | lambda *args, **kwargs: [str(tool_root)], |
| 270 | ) |
| 271 | monkeypatch.setattr( |
| 272 | mcp_handler.MCPConfig, |
| 273 | "get_for_agent", |
| 274 | lambda agent: _NoMCPTools(), |
| 275 | ) |
| 276 | monkeypatch.setattr( |
| 277 | tool_policy, |
| 278 | "get_policy", |
| 279 | lambda agent: { |
| 280 | "mode": "inherit", |
| 281 | "default": "allow", |
| 282 | "allowed": [], |
| 283 | "blocked": [], |
| 284 | }, |
| 285 | ) |
| 286 | |
| 287 | catalog = tool_policy.get_tool_catalog(_Agent(tmp_path)) |
| 288 | |
| 289 | assert [item["name"] for item in catalog] == ["code_execution_remote", "shell"] |
| 290 | |
| 291 | |
| 292 | def test_mcp_catalog_labels_include_humanized_server_and_tool( |
| 293 | monkeypatch, tmp_path: Path |
| 294 | ) -> None: |
| 295 | class MCPTools: |
| 296 | def get_tools(self): |
| 297 | return [ |
| 298 | { |
| 299 | "deep_wiki.ask_question": { |
| 300 | "name": "ask_question", |
| 301 | "description": "Ask DeepWiki", |
| 302 | "server": "deep_wiki", |
| 303 | } |
| 304 | } |
| 305 | ] |
| 306 | |
| 307 | monkeypatch.setattr(tool_policy.subagents, "get_paths", lambda *args, **kwargs: []) |
| 308 | monkeypatch.setattr(mcp_handler.MCPConfig, "get_for_agent", lambda agent: MCPTools()) |
| 309 | monkeypatch.setattr( |
| 310 | tool_policy, |
| 311 | "get_policy", |
| 312 | lambda agent: { |
| 313 | "mode": "inherit", |
| 314 | "default": "allow", |
| 315 | "allowed": [], |
| 316 | "blocked": [], |
| 317 | }, |
| 318 | ) |
| 319 | |
| 320 | assert tool_policy.get_tool_catalog(_Agent(tmp_path)) == [ |
| 321 | { |
| 322 | "id": "mcp:deep_wiki:ask_question", |
| 323 | "name": "deep_wiki.ask_question", |
| 324 | "label": "Deep Wiki · Ask Question", |
| 325 | "description": "Ask DeepWiki", |
| 326 | "origin": "MCP · deep_wiki", |
| 327 | "available": True, |
| 328 | } |
| 329 | ] |
| 330 | |
| 331 | |
| 332 | @pytest.mark.asyncio |
| 333 | async def test_skills_catalog_prompt_is_absent_when_skills_tool_is_blocked( |
| 334 | monkeypatch, tmp_path: Path |
| 335 | ) -> None: |
| 336 | monkeypatch.setattr(tool_policy.subagents, "get_paths", lambda *args, **kwargs: []) |
| 337 | monkeypatch.setattr( |
| 338 | tool_policy, |
| 339 | "get_policy", |
| 340 | lambda agent: _custom_policy(default="allow", blocked=["local:skills_tool"]), |
| 341 | ) |
| 342 | monkeypatch.setattr( |
| 343 | _13_skills_prompt.skills_helper, |
| 344 | "list_skills", |
| 345 | lambda **kwargs: pytest.fail("blocked skill discovery ran"), |
| 346 | ) |
| 347 | |
| 348 | assert await _13_skills_prompt.build_prompt(_Agent(tmp_path)) == "" |
| 349 | |
| 350 | |
| 351 | def test_tool_prompt_description_skips_fenced_examples() -> None: |
| 352 | prompt = """### example |
| 353 | ~~~json |
| 354 | {"tool_name":"example","tool_args":{}} |
| 355 | ~~~ |
| 356 | Visible summary |
| 357 | """ |
| 358 | |
| 359 | assert tool_policy.tool_prompt_description(prompt, "example") == "Visible summary" |
| 360 | |
| 361 | |
| 362 | def test_tool_prompt_description_prefers_declared_summary() -> None: |
| 363 | prompt = """## tools |
| 364 | - `memory_load`: search stored memories by meaning and metadata |
| 365 | args: `query`, optional `limit` |
| 366 | """ |
| 367 | |
| 368 | assert ( |
| 369 | tool_policy.tool_prompt_description(prompt, "memory_load") |
| 370 | == "search stored memories by meaning and metadata" |
| 371 | ) |
| 372 | |
| 373 | |
| 374 | def test_prompt_filter_removes_complete_blocked_json_example( |
| 375 | monkeypatch, tmp_path: Path |
| 376 | ) -> None: |
| 377 | monkeypatch.setattr( |
| 378 | tool_policy, |
| 379 | "_policy_tool_names", |
| 380 | lambda agent: {"memory_load", "memory_save"}, |
| 381 | ) |
| 382 | monkeypatch.setattr(tool_policy.subagents, "get_paths", lambda *args, **kwargs: []) |
| 383 | monkeypatch.setattr( |
| 384 | tool_policy, |
| 385 | "get_policy", |
| 386 | lambda agent: _custom_policy( |
| 387 | default="allow", blocked=["local:memory_load"] |
| 388 | ), |
| 389 | ) |
| 390 | prompt = """## memory tools |
| 391 | - `memory_load`: load memory |
| 392 | - `memory_save`: save memory |
| 393 | ~~~json |
| 394 | { |
| 395 | "tool_name": "memory_load", |
| 396 | "tool_args": {"query": "blocked example"} |
| 397 | } |
| 398 | ~~~ |
| 399 | ~~~json |
| 400 | { |
| 401 | "tool_name": "memory_save", |
| 402 | "tool_args": {"text": "allowed example"} |
| 403 | } |
| 404 | ~~~ |
| 405 | """ |
| 406 | |
| 407 | filtered = tool_policy.filter_tool_prompt( |
| 408 | _Agent(tmp_path), "agent.system.tool.memory.md", prompt |
| 409 | ) |
| 410 | |
| 411 | assert "blocked example" not in filtered |
| 412 | assert "memory_load" not in filtered |
| 413 | assert "allowed example" in filtered |
| 414 | assert filtered.count("~~~json") == 1 |
| 415 | assert filtered.count("~~~") == 2 |
| 416 | |
| 417 | |
| 418 | def test_plugin_tool_identity_uses_canonical_plugin_roots( |
| 419 | monkeypatch, tmp_path: Path |
| 420 | ) -> None: |
| 421 | plugin_root = tmp_path / "plugins" / "_example" |
| 422 | plugin_tool = plugin_root / "tools" / "actual.py" |
| 423 | plugin_tool.parent.mkdir(parents=True) |
| 424 | plugin_tool.write_text("class Actual: pass\n", encoding="utf-8") |
| 425 | monkeypatch.setattr( |
| 426 | tool_policy.plugins, |
| 427 | "get_plugin_roots", |
| 428 | lambda: [str(tmp_path / "usr" / "plugins"), str(tmp_path / "plugins")], |
| 429 | ) |
| 430 | monkeypatch.setattr( |
| 431 | tool_policy, |
| 432 | "get_policy", |
| 433 | lambda agent: { |
| 434 | "mode": "inherit", |
| 435 | "default": "allow", |
| 436 | "allowed": [], |
| 437 | "blocked": [], |
| 438 | }, |
| 439 | ) |
| 440 | agent = _Agent(tmp_path) |
| 441 | |
| 442 | monkeypatch.setattr( |
| 443 | tool_policy.subagents, |
| 444 | "get_paths", |
| 445 | lambda *args, **kwargs: [str(plugin_tool)], |
| 446 | ) |
| 447 | assert tool_policy.resolve_tool(agent, "actual").tool_id == "plugin:_example:actual" |
| 448 | |
| 449 | lookalike = tmp_path / "work" / "plugins" / "_example" / "tools" / "actual.py" |
| 450 | lookalike.parent.mkdir(parents=True) |
| 451 | lookalike.write_text("class Actual: pass\n", encoding="utf-8") |
| 452 | monkeypatch.setattr( |
| 453 | tool_policy.subagents, |
| 454 | "get_paths", |
| 455 | lambda *args, **kwargs: [str(lookalike)], |
| 456 | ) |
| 457 | assert tool_policy.resolve_tool(agent, "actual").tool_id == "local:actual" |
| 458 | |
| 459 | |
| 460 | def test_dotted_local_tool_keeps_local_identity_at_execution_gate( |
| 461 | monkeypatch, tmp_path: Path |
| 462 | ) -> None: |
| 463 | tool_path = tmp_path / "docs.read.py" |
| 464 | tool_path.write_text("class Tool: pass\n", encoding="utf-8") |
| 465 | monkeypatch.setattr( |
| 466 | tool_policy.subagents, |
| 467 | "get_paths", |
| 468 | lambda *args, **kwargs: [str(tool_path)], |
| 469 | ) |
| 470 | monkeypatch.setattr( |
| 471 | mcp_handler.MCPConfig, |
| 472 | "get_for_agent", |
| 473 | lambda agent: _NoMCPTools(), |
| 474 | ) |
| 475 | monkeypatch.setattr( |
| 476 | tool_policy, |
| 477 | "get_policy", |
| 478 | lambda agent: _custom_policy( |
| 479 | default="allow", blocked=["local:docs.read"] |
| 480 | ), |
| 481 | ) |
| 482 | agent = _Agent(tmp_path) |
| 483 | |
| 484 | with pytest.raises(RepairableException, match='Tool "docs.read" is blocked'): |
| 485 | tool_policy.ensure_tool_allowed(agent, "docs.read") |
| 486 | |
| 487 | |
| 488 | def test_legacy_response_and_vision_policy_ids_stay_out_of_catalog( |
| 489 | monkeypatch, tmp_path: Path |
| 490 | ) -> None: |
| 491 | monkeypatch.setattr(tool_policy.subagents, "get_paths", lambda *args, **kwargs: []) |
| 492 | monkeypatch.setattr( |
| 493 | mcp_handler.MCPConfig, |
| 494 | "get_for_agent", |
| 495 | lambda agent: _NoMCPTools(), |
| 496 | ) |
| 497 | monkeypatch.setattr( |
| 498 | tool_policy, |
| 499 | "get_policy", |
| 500 | lambda agent: _custom_policy( |
| 501 | default="allow", |
| 502 | blocked=[ |
| 503 | "response", |
| 504 | "local:response", |
| 505 | "plugin:legacy:response", |
| 506 | "vision_load", |
| 507 | "local:vision_load", |
| 508 | "plugin:legacy:vision_load", |
| 509 | ], |
| 510 | ), |
| 511 | ) |
| 512 | |
| 513 | assert tool_policy.get_tool_catalog(_Agent(tmp_path)) == [] |
| 514 | |
| 515 | |
| 516 | @pytest.mark.asyncio |
| 517 | async def test_vision_tool_follows_chat_config_not_profile_policy( |
| 518 | monkeypatch, tmp_path: Path |
| 519 | ) -> None: |
| 520 | _write_prompt(tmp_path, "agent.system.tools.md", "TOOLS\n{{tools}}") |
| 521 | _write_prompt( |
| 522 | tmp_path, |
| 523 | "agent.system.tools_vision.md", |
| 524 | '### vision_load\nload images\n{"tool_name":"vision_load","tool_args":{"paths":[]}}', |
| 525 | ) |
| 526 | monkeypatch.setattr(tool_policy.subagents, "get_paths", _prompt_paths(tmp_path)) |
| 527 | monkeypatch.setattr( |
| 528 | "plugins._model_config.helpers.model_config.get_chat_model_config", |
| 529 | lambda agent: {"vision": True}, |
| 530 | ) |
| 531 | monkeypatch.setattr( |
| 532 | "plugins._model_config.helpers.model_config.get_vision_model_config", |
| 533 | lambda agent: {}, |
| 534 | ) |
| 535 | monkeypatch.setattr( |
| 536 | tool_policy, |
| 537 | "get_policy", |
| 538 | lambda agent: _custom_policy( |
| 539 | default="block", blocked=["local:vision_load"] |
| 540 | ), |
| 541 | ) |
| 542 | monkeypatch.setattr(responses_tools, "_mcp_tools", lambda agent: []) |
| 543 | agent = _Agent(tmp_path) |
| 544 | |
| 545 | prompt = await _11_tools_prompt.build_prompt(agent) |
| 546 | schemas, _name_map = responses_tools.build_responses_function_tools(agent) |
| 547 | |
| 548 | assert "vision_load" in prompt |
| 549 | assert [schema["name"] for schema in schemas] == ["vision_load"] |
| 550 | assert tool_policy.resolve_tool(agent, "vision_load").source == "runtime-config" |
| 551 | |
| 552 | |
| 553 | @pytest.mark.asyncio |
| 554 | async def test_active_vision_model_uses_canonical_vision_prompt( |
| 555 | monkeypatch, tmp_path: Path |
| 556 | ) -> None: |
| 557 | _write_prompt(tmp_path, "agent.system.tools.md", "TOOLS\n{{tools}}") |
| 558 | _write_prompt( |
| 559 | tmp_path, |
| 560 | "agent.system.tools_vision.md", |
| 561 | "### vision_load\ncanonical vision\nargs: `paths`, `query`", |
| 562 | ) |
| 563 | monkeypatch.setattr(tool_policy.subagents, "get_paths", _prompt_paths(tmp_path)) |
| 564 | monkeypatch.setattr( |
| 565 | "plugins._model_config.helpers.model_config.get_chat_model_config", |
| 566 | lambda agent: {"vision": False}, |
| 567 | ) |
| 568 | monkeypatch.setattr( |
| 569 | "plugins._model_config.helpers.model_config.get_vision_model_config", |
| 570 | lambda agent: {"provider": "test", "name": "vision"}, |
| 571 | ) |
| 572 | monkeypatch.setattr(responses_tools, "_mcp_tools", lambda agent: []) |
| 573 | agent = _Agent(tmp_path) |
| 574 | |
| 575 | prompt = await _11_tools_prompt.build_prompt(agent) |
| 576 | schemas, _name_map = responses_tools.build_responses_function_tools(agent) |
| 577 | |
| 578 | assert prompt.count("canonical vision") == 1 |
| 579 | assert schemas[0]["name"] == "vision_load" |
| 580 | assert schemas[0]["description"] == "canonical vision" |
| 581 | |
| 582 | |
| 583 | def test_vision_prompt_stays_route_agnostic_and_batches_paths() -> None: |
| 584 | source = ( |
| 585 | Path(__file__).resolve().parents[1] |
| 586 | / "prompts" |
| 587 | / "agent.system.tools_vision.md" |
| 588 | ).read_text(encoding="utf-8") |
| 589 | schema = responses_tools._schema_from_prompt(source) |
| 590 | |
| 591 | assert schema == { |
| 592 | "type": "object", |
| 593 | "properties": {}, |
| 594 | "additionalProperties": True, |
| 595 | } |
| 596 | assert "load all relevant images in one call" in source |
| 597 | assert "Input schema for tool_args" not in source |
| 598 | assert "Vision Model" not in source |
| 599 | assert "sidecar" not in source.lower() |
| 600 | assert "optional `query`" in source |
| 601 | |
| 602 | |
| 603 | def test_vision_framework_prompt_renders_optional_query() -> None: |
| 604 | prompt_dir = str(Path(__file__).resolve().parents[1] / "prompts") |
| 605 | |
| 606 | without_query = files.read_prompt_file( |
| 607 | "fw.vision_load.md", |
| 608 | _directories=[prompt_dir], |
| 609 | request="Review these screenshots.", |
| 610 | query="", |
| 611 | ) |
| 612 | with_query = files.read_prompt_file( |
| 613 | "fw.vision_load.md", |
| 614 | _directories=[prompt_dir], |
| 615 | request="Review these screenshots.", |
| 616 | query="Compare the error banners.", |
| 617 | ) |
| 618 | |
| 619 | assert "Visual query:" not in without_query |
| 620 | assert "Current request:\nReview these screenshots." in with_query |
| 621 | assert "Visual query:\nCompare the error banners." in with_query |
| 622 | |
| 623 | |
| 624 | def test_mcp_prompt_and_native_schema_omit_blocked_tool( |
| 625 | monkeypatch, tmp_path: Path |
| 626 | ) -> None: |
| 627 | class Server: |
| 628 | name = "docs" |
| 629 | description = "Documentation" |
| 630 | |
| 631 | def get_tools(self): |
| 632 | return [ |
| 633 | { |
| 634 | "name": "read", |
| 635 | "description": "Read docs", |
| 636 | "input_schema": {"type": "object"}, |
| 637 | }, |
| 638 | { |
| 639 | "name": "write", |
| 640 | "description": "Write docs", |
| 641 | "input_schema": {"type": "object"}, |
| 642 | }, |
| 643 | ] |
| 644 | |
| 645 | config = mcp_handler.MCPConfig(servers_list=[]) |
| 646 | config.servers = [Server()] |
| 647 | agent = _Agent(tmp_path) |
| 648 | monkeypatch.setattr( |
| 649 | tool_policy, |
| 650 | "get_policy", |
| 651 | lambda agent: _custom_policy( |
| 652 | default="allow", blocked=["mcp:docs:write"] |
| 653 | ), |
| 654 | ) |
| 655 | monkeypatch.setattr( |
| 656 | responses_tools, |
| 657 | "_mcp_tools", |
| 658 | lambda agent: [ |
| 659 | ("docs.read", Server().get_tools()[0]), |
| 660 | ("docs.write", Server().get_tools()[1]), |
| 661 | ], |
| 662 | ) |
| 663 | monkeypatch.setattr(tool_policy.subagents, "get_paths", lambda *args: []) |
| 664 | monkeypatch.setattr(responses_tools, "_vision_tool_prompt", lambda agent: "") |
| 665 | |
| 666 | prompt = config.get_tools_prompt(agent=agent) |
| 667 | schemas, name_map = responses_tools.build_responses_function_tools(agent) |
| 668 | |
| 669 | assert "docs.read" in prompt |
| 670 | assert "docs.write" not in prompt |
| 671 | assert len(schemas) == 1 |
| 672 | assert name_map[schemas[0]["name"]] == "docs.read" |
| 673 | |
| 674 | |
| 675 | @pytest.mark.asyncio |
| 676 | async def test_local_execution_gate_returns_stable_profile_error( |
| 677 | monkeypatch, tmp_path: Path |
| 678 | ) -> None: |
| 679 | agent = _Agent(tmp_path, profile="researcher") |
| 680 | monkeypatch.setattr(tool_policy.subagents, "get_paths", lambda *args, **kwargs: []) |
| 681 | monkeypatch.setattr( |
| 682 | tool_policy, |
| 683 | "get_policy", |
| 684 | lambda agent: _custom_policy(default="block"), |
| 685 | ) |
| 686 | |
| 687 | with pytest.raises( |
| 688 | RepairableException, |
| 689 | match='Tool "shell" is blocked for agent profile "researcher"', |
| 690 | ): |
| 691 | await EnforceToolPolicy(agent).execute(tool_name="shell") |
| 692 | |
| 693 | |
| 694 | @pytest.mark.asyncio |
| 695 | async def test_mcp_invocation_rechecks_policy_before_server_call( |
| 696 | monkeypatch, tmp_path: Path |
| 697 | ) -> None: |
| 698 | agent = _Agent(tmp_path, profile="researcher") |
| 699 | called = False |
| 700 | |
| 701 | class Config: |
| 702 | async def call_tool(self, name, kwargs): |
| 703 | nonlocal called |
| 704 | called = True |
| 705 | raise AssertionError("blocked MCP call reached the server") |
| 706 | |
| 707 | monkeypatch.setattr(mcp_handler.MCPConfig, "get_for_agent", lambda agent: Config()) |
| 708 | monkeypatch.setattr( |
| 709 | tool_policy, |
| 710 | "get_policy", |
| 711 | lambda agent: _custom_policy( |
| 712 | default="allow", blocked=["mcp:docs:write"] |
| 713 | ), |
| 714 | ) |
| 715 | tool = mcp_handler.MCPTool( |
| 716 | agent=agent, |
| 717 | name="docs.write", |
| 718 | method=None, |
| 719 | args={}, |
| 720 | message="", |
| 721 | loop_data=None, |
| 722 | ) |
| 723 | |
| 724 | with pytest.raises(RepairableException, match='Tool "docs.write" is blocked'): |
| 725 | await tool.execute() |
| 726 | assert called is False |
| 727 | |
| 728 | |
| 729 | @pytest.mark.asyncio |
| 730 | async def test_delegated_agent_uses_its_own_profile_policy_at_execution_gate( |
| 731 | monkeypatch, tmp_path: Path |
| 732 | ) -> None: |
| 733 | parent = _Agent(tmp_path, profile="agent0") |
| 734 | child = _Agent(tmp_path, profile="researcher") |
| 735 | monkeypatch.setattr(tool_policy.subagents, "get_paths", lambda *args, **kwargs: []) |
| 736 | |
| 737 | def policy_for_profile(agent): |
| 738 | if agent.config.profile == "researcher": |
| 739 | return _custom_policy(default="block") |
| 740 | return {"mode": "inherit"} |
| 741 | |
| 742 | monkeypatch.setattr(tool_policy, "get_policy", policy_for_profile) |
| 743 | |
| 744 | assert tool_policy.resolve_tool(parent, "shell").allowed is True |
| 745 | assert tool_policy.resolve_tool(child, "shell").allowed is False |
| 746 | await EnforceToolPolicy(parent).execute(tool_name="shell") |
| 747 | with pytest.raises( |
| 748 | RepairableException, |
| 749 | match='Tool "shell" is blocked for agent profile "researcher"', |
| 750 | ): |
| 751 | await EnforceToolPolicy(child).execute(tool_name="shell") |
| 752 | |
| 753 | |
| 754 | def test_project_policy_precedes_profile_policy( |
| 755 | monkeypatch: pytest.MonkeyPatch, |
| 756 | tmp_path: Path, |
| 757 | ) -> None: |
| 758 | class ProjectContext: |
| 759 | def get_data(self, key: str, recursive: bool = True): |
| 760 | return "demo" if key == "project" else None |
| 761 | |
| 762 | monkeypatch.setattr(tool_policy.files, "_base_dir", str(tmp_path)) |
| 763 | monkeypatch.setattr(tool_policy.subagents, "get_paths", lambda *args, **kwargs: []) |
| 764 | monkeypatch.setattr( |
| 765 | tool_policy.plugins, |
| 766 | "call_plugin_hook", |
| 767 | lambda _plugin, _hook, default=None, **_kwargs: default, |
| 768 | ) |
| 769 | agent = _Agent(tmp_path) |
| 770 | agent.context = ProjectContext() |
| 771 | |
| 772 | tool_policy.plugins.save_plugin_config( |
| 773 | tool_policy.PLUGIN_NAME, |
| 774 | "", |
| 775 | "researcher", |
| 776 | _custom_policy(default="block"), |
| 777 | ) |
| 778 | tool_policy.plugins.save_plugin_config( |
| 779 | tool_policy.PLUGIN_NAME, |
| 780 | "demo", |
| 781 | "", |
| 782 | _custom_policy(default="allow"), |
| 783 | ) |
| 784 | tool_policy.plugins.save_plugin_config( |
| 785 | tool_policy.PLUGIN_NAME, |
| 786 | "demo", |
| 787 | "researcher", |
| 788 | _custom_policy(default="allow", blocked=["local:shell"]), |
| 789 | ) |
| 790 | profile_path = Path( |
| 791 | tool_policy.plugins.determine_plugin_asset_path( |
| 792 | tool_policy.PLUGIN_NAME, |
| 793 | "", |
| 794 | "researcher", |
| 795 | tool_policy.plugins.CONFIG_FILE_NAME, |
| 796 | ) |
| 797 | ) |
| 798 | project_path = Path( |
| 799 | tool_policy.plugins.determine_plugin_asset_path( |
| 800 | tool_policy.PLUGIN_NAME, |
| 801 | "demo", |
| 802 | "", |
| 803 | tool_policy.plugins.CONFIG_FILE_NAME, |
| 804 | ) |
| 805 | ) |
| 806 | project_profile_path = Path( |
| 807 | tool_policy.plugins.determine_plugin_asset_path( |
| 808 | tool_policy.PLUGIN_NAME, |
| 809 | "demo", |
| 810 | "researcher", |
| 811 | tool_policy.plugins.CONFIG_FILE_NAME, |
| 812 | ) |
| 813 | ) |
| 814 | |
| 815 | decision = tool_policy.resolve_tool(agent, "shell") |
| 816 | assert decision.allowed is False |
| 817 | assert decision.source == "scoped-policy" |
| 818 | |
| 819 | project_profile_path.write_text('{"manual": true}\n', encoding="utf-8") |
| 820 | decision = tool_policy.resolve_tool(agent, "shell") |
| 821 | assert decision.allowed is True |
| 822 | assert decision.source == "scoped-default" |
| 823 | assert project_profile_path.read_text(encoding="utf-8") == '{"manual": true}\n' |
| 824 | |
| 825 | project_profile_path.write_text( |
| 826 | '{"manual": true, "mode": "inherit", "default": "block", ' |
| 827 | '"allowed": [], "blocked": ["local:shell"]}\n', |
| 828 | encoding="utf-8", |
| 829 | ) |
| 830 | decision = tool_policy.resolve_tool(agent, "shell") |
| 831 | assert decision.allowed is True |
| 832 | assert decision.source == "scoped-default" |
| 833 | |
| 834 | project_path.unlink() |
| 835 | decision = tool_policy.resolve_tool(agent, "shell") |
| 836 | assert decision.allowed is False |
| 837 | assert decision.source == "scoped-default" |
| 838 | assert profile_path.is_file() |
| 839 | |
| 840 | |
| 841 | def test_unknown_policy_ids_are_retained_as_unavailable( |
| 842 | monkeypatch, tmp_path: Path |
| 843 | ) -> None: |
| 844 | agent = _Agent(tmp_path) |
| 845 | monkeypatch.setattr(tool_policy.subagents, "get_paths", lambda *args, **kwargs: []) |
| 846 | monkeypatch.setattr( |
| 847 | mcp_handler.MCPConfig, |
| 848 | "get_for_agent", |
| 849 | lambda agent: _NoMCPTools(), |
| 850 | ) |
| 851 | monkeypatch.setattr( |
| 852 | tool_policy, |
| 853 | "get_policy", |
| 854 | lambda agent: _custom_policy( |
| 855 | default="allow", blocked=["plugin:missing:ghost"] |
| 856 | ), |
| 857 | ) |
| 858 | |
| 859 | catalog = tool_policy.get_tool_catalog(agent) |
| 860 | |
| 861 | assert catalog == [ |
| 862 | { |
| 863 | "id": "plugin:missing:ghost", |
| 864 | "name": "ghost", |
| 865 | "label": "Ghost", |
| 866 | "description": "", |
| 867 | "origin": "Unavailable", |
| 868 | "available": False, |
| 869 | } |
| 870 | ] |