diff --git a/agent_platform/runtime/loop.py b/agent_platform/runtime/loop.py index c0b9e66..4729044 100644 --- a/agent_platform/runtime/loop.py +++ b/agent_platform/runtime/loop.py @@ -33,6 +33,8 @@ relevant verification, and report concrete results. Use tools instead of inventi command output. A claim that a file was created, changed, executed, or tested must be backed by a successful tool result from this run. For scripts, reports, code, and other deliverables, create real files under /workspace, inspect them after writing, run relevant checks, and cite their exact paths. +For a code-and-report task, follow this order: write a real source file, read it, run that exact source +file with a complete command, write the report using only the measured output, then read the report. Write source files with real line breaks rather than literal `\\n` escape sequences. If any command fails, inspect the failure, repair the underlying problem, and successfully rerun the relevant check; never replace failed output with invented numbers, placeholders, or a hand-written success report. @@ -113,6 +115,32 @@ READ_ONLY_EXEC_COMMAND = re.compile( r"^(?:cat|cut|diff|find|git\s+(?:diff|show|status)|grep|head|jq|ls|pwd|rg|sed\s+-n|stat|tail|wc)\b", re.IGNORECASE, ) +PROCESS_REQUEST = re.compile( + r"(服务器|服务|后台|常驻|网站|网页|\b(?:server|service|daemon|background|web\s*app)\b)", + re.IGNORECASE, +) +GIT_REQUEST = re.compile( + r"(仓库|提交|分支|\b(?:git|repo(?:sitory)?|commit|branch|pull request)\b)", + re.IGNORECASE, +) +MEMORY_REQUEST = re.compile( + r"(记住|记忆|偏好|忘记|\b(?:remember|memory|preference|forget)\b)", + re.IGNORECASE, +) +DELEGATION_REQUEST = re.compile( + r"(多\s*(?:agent|代理)|子代理|分工|并行调研|大型|架构|重构|迁移|" + r"\b(?:multi[- ]agent|sub[- ]agent|delegate|parallel research|architecture|migration|refactor)\b)", + re.IGNORECASE, +) +CORE_WORK_TOOLS = { + "workspace_status", + "list_files", + "read_file", + "search_files", + "write_file", + "apply_patch", + "exec", +} def latest_user_text(messages: list[dict[str, Any]]) -> str: @@ -300,6 +328,26 @@ def is_substantive_execution(name: str, arguments: dict[str, Any]) -> bool: return bool(command and not READ_ONLY_EXEC_COMMAND.match(command)) +def select_tool_specs( + specs: list[dict[str, Any]], + request_text: str, + *, + depth: int, +) -> list[dict[str, Any]]: + allowed = set(CORE_WORK_TOOLS) + if depth == 0: + allowed.add("update_plan") + if PROCESS_REQUEST.search(request_text): + allowed.update({"start_process", "poll_process", "cancel_process"}) + if GIT_REQUEST.search(request_text): + allowed.update({"git_status", "git_diff"}) + if MEMORY_REQUEST.search(request_text): + allowed.update({"remember", "recall_memory", "forget_memory"}) + if depth == 0 and DELEGATION_REQUEST.search(request_text): + allowed.add("delegate_task") + return [spec for spec in specs if str((spec.get("function") or {}).get("name", "")) in allowed] + + @dataclass(slots=True) class RunRecorder: store: RuntimeStore @@ -403,8 +451,12 @@ class AgentLoop: read_only: bool, ) -> str: max_iterations = min(spec.max_iterations, 8 if depth else spec.max_iterations) - available_specs = self.tools.specs(read_only=read_only, allow_delegate=depth == 0) request_text = latest_user_text(messages) + available_specs = select_tool_specs( + self.tools.specs(read_only=read_only, allow_delegate=depth == 0), + request_text, + depth=depth, + ) artifact_required = depth == 0 and bool(ARTIFACT_REQUEST.search(request_text)) action_required = depth == 0 and bool(ACTION_REQUEST.search(request_text)) execution_required = depth == 0 and bool(EXECUTION_REQUEST.search(request_text)) diff --git a/agent_platform/runtime/tools.py b/agent_platform/runtime/tools.py index 23a79a4..413b8ef 100644 --- a/agent_platform/runtime/tools.py +++ b/agent_platform/runtime/tools.py @@ -92,7 +92,7 @@ TOOL_METADATA: dict[str, ToolMetadata] = { ), "write_file": ToolMetadata( "write_file", - "Write the complete UTF-8 contents of a workspace file.", + "Write complete UTF-8 file contents. Use actual line breaks between source lines, not literal \\\\n text.", object_schema( { "path": {"type": "string", "maxLength": 4096}, @@ -115,7 +115,8 @@ TOOL_METADATA: dict[str, ToolMetadata] = { ), "exec": ToolMetadata( "exec", - "Run a shell command in the isolated workspace.", + "Run a complete non-interactive shell command. Name the script, pass -c code, or invoke a test; " + "bare python, node, or shells are rejected.", object_schema( { "command": {"type": "string"}, diff --git a/tests/test_agent_loop.py b/tests/test_agent_loop.py index cb0b0ca..990e6ec 100644 --- a/tests/test_agent_loop.py +++ b/tests/test_agent_loop.py @@ -11,6 +11,7 @@ from agent_platform.runtime.loop import ( AgentLoop, is_substantive_execution, normalize_write_file_content, + select_tool_specs, tool_event_details, ) from agent_platform.runtime.tools import TOOL_METADATA @@ -45,6 +46,34 @@ def test_read_only_shell_commands_do_not_satisfy_execution_evidence() -> None: assert is_substantive_execution("start_process", {"command": "python3 server.py"}) +def test_simple_script_task_receives_a_focused_tool_menu() -> None: + specs = [metadata.openai_spec() for metadata in TOOL_METADATA.values()] + selected = select_tool_specs(specs, "写脚本对比排序算法，给一个报告", depth=0) + names = {spec["function"]["name"] for spec in selected} + assert names == { + "workspace_status", + "list_files", + "read_file", + "search_files", + "write_file", + "apply_patch", + "exec", + "update_plan", + } + + +def test_complex_requests_enable_only_relevant_optional_tools() -> None: + specs = [metadata.openai_spec() for metadata in TOOL_METADATA.values()] + selected = select_tool_specs( + specs, + "重构 Git 仓库里的 web server，并记住这个偏好，分工并行调研", + depth=0, + ) + names = {spec["function"]["name"] for spec in selected} + assert {"delegate_task", "git_status", "git_diff", "start_process", "poll_process"} <= names + assert {"remember", "recall_memory", "forget_memory"} <= names + + class ScriptedProvider: def __init__(self, responses: list[dict[str, Any]]) -> None: self.responses = responses