diff --git a/.github/workflows/release.yml b/.github/workflows/release.yml new file mode 100644 index 0000000..d923427 --- /dev/null +++ b/.github/workflows/release.yml @@ -0,0 +1,91 @@ +name: Release + +# Triggered when a version tag is pushed, e.g.: +# git tag v0.7.5.0 && git push origin v0.7.5.0 +on: + push: + tags: + - "v*" + +permissions: + contents: read + +jobs: + # 1. Build the sdist + wheel and verify the tag matches pyproject version. + build: + runs-on: ubuntu-latest + steps: + - name: Checkout code + uses: actions/checkout@v4 + + - name: Install uv + uses: astral-sh/setup-uv@v3 + with: + version: "latest" + + - name: Set up Python + uses: actions/setup-python@v4 + with: + python-version: "3.13" + + - name: Verify tag matches pyproject version + run: | + TAG="${GITHUB_REF_NAME#v}" + PKG=$(uv version --short 2>/dev/null || python -c "import tomllib;print(tomllib.load(open('pyproject.toml','rb'))['project']['version'])") + echo "Tag version: $TAG" + echo "Package version: $PKG" + if [ "$TAG" != "$PKG" ]; then + echo "::error::Tag ($TAG) does not match pyproject.toml version ($PKG)." + exit 1 + fi + + - name: Build sdist and wheel + run: uv build + + - name: Check distribution metadata + run: uvx twine check dist/* + + - name: Upload build artifacts + uses: actions/upload-artifact@v4 + with: + name: dist + path: dist/ + + # 2. Publish to PyPI via Trusted Publishing (OIDC, no API token needed). + pypi: + needs: build + runs-on: ubuntu-latest + environment: + name: pypi + url: https://pypi.org/project/10xscale-agentflow/ + permissions: + id-token: write # required for trusted publishing + steps: + - name: Download build artifacts + uses: actions/download-artifact@v4 + with: + name: dist + path: dist/ + + - name: Publish to PyPI + uses: pypa/gh-action-pypi-publish@release/v1 + + # 3. Create the GitHub Release with auto-generated notes and attach artifacts. + github-release: + needs: pypi + runs-on: ubuntu-latest + permissions: + contents: write # required to create a release + steps: + - name: Download build artifacts + uses: actions/download-artifact@v4 + with: + name: dist + path: dist/ + + - name: Create GitHub Release + uses: softprops/action-gh-release@v2 + with: + generate_release_notes: true + files: dist/* + fail_on_unmatched_files: true diff --git a/agentflow/core/graph/utils/invoke_handler.py b/agentflow/core/graph/utils/invoke_handler.py index 47645e3..309e251 100644 --- a/agentflow/core/graph/utils/invoke_handler.py +++ b/agentflow/core/graph/utils/invoke_handler.py @@ -11,6 +11,7 @@ from agentflow.core.graph.edge import Edge from agentflow.core.graph.node import Node from agentflow.core.graph.utils.utils import ( + calculate_token_usage, call_realtime_sync, get_next_node, load_or_create_state, @@ -479,8 +480,12 @@ async def invoke( final_state, messages = await self._execute_graph(state, config) logger.info("Graph execution completed with %d final messages", len(messages)) + # Calculate token usage + token_usage = calculate_token_usage(messages) + event.event_type = EventType.END event.metadata["status"] = "Graph execution completed" + event.metadata.update(token_usage) event.data["state"] = final_state.model_dump() event.data["messages"] = [m.model_dump() for m in messages] if messages else [] publish_event(event) @@ -489,6 +494,7 @@ async def invoke( final_state, messages, response_granularity, + token_usage=token_usage, ) except Exception as e: logger.exception("Graph execution failed: %s", e) diff --git a/agentflow/core/graph/utils/stream_handler.py b/agentflow/core/graph/utils/stream_handler.py index bdb5710..ee8cafc 100644 --- a/agentflow/core/graph/utils/stream_handler.py +++ b/agentflow/core/graph/utils/stream_handler.py @@ -39,6 +39,7 @@ InterruptConfigMixin, ) from .utils import ( + calculate_token_usage, call_realtime_sync, get_next_node, load_or_create_state, @@ -589,6 +590,7 @@ async def _execute_graph( # noqa: PLR0912, PLR0915 event.metadata["is_context_trimmed"] = is_context_trimmed publish_event(event) + # Include messages list for token calculation in stream method yield StreamChunk( event=StreamEvent.UPDATES, state=state, @@ -599,6 +601,8 @@ async def _execute_graph( # noqa: PLR0912, PLR0915 "max_steps": max_steps, "is_context_trimmed": is_context_trimmed, "reason": "Graph execution completed successfully", + # Internal: messages from current run for token calculation + "_messages": messages, }, thread_id=config.get("thread_id"), run_id=config.get("run_id"), @@ -747,8 +751,19 @@ async def stream( logger.debug("Beginning graph execution") result = self._execute_graph(state, input_data, config) + # Track messages from current run for token calculation + current_run_messages = [] + # Stream results based on response granularity async for chunk in result: + # Extract messages from final completion chunk (internal use only) + if ( + chunk.event == StreamEvent.UPDATES + and chunk.data + and chunk.data.get("status") == "graph_invoked" + ): + current_run_messages = chunk.data.pop("_messages", []) + match response_granularity: case ResponseGranularity.FULL: yield chunk @@ -763,6 +778,9 @@ async def stream( time_taken = time.time() - start_time logger.info("Graph execution finished in %.2f seconds", time_taken) + # Calculate token usage from current run messages only + token_usage = calculate_token_usage(current_run_messages) + event.event_type = EventType.END event.metadata.update( { @@ -772,19 +790,22 @@ async def stream( "current_node": state.execution_meta.current_node, "is_interrupted": state.is_interrupted(), "total_messages": len(state.context) if state.context else 0, + **token_usage, } ) publish_event(event) - yield StreamChunk( - event=StreamEvent.UPDATES, - state=state, - data={ - "status": "graph_invoked", - "reason": "Graph execution finished", - "time_taken": time_taken, - "is_interrupted": state.is_interrupted(), - "total_messages": len(state.context) if state.context else 0, - }, - thread_id=config.get("thread_id"), - run_id=config.get("run_id"), - ) + if response_granularity == ResponseGranularity.FULL: + yield StreamChunk( + event=StreamEvent.UPDATES, + state=state, + data={ + "status": "graph_invoked", + "reason": "Graph execution finished", + "time_taken": time_taken, + "is_interrupted": state.is_interrupted(), + "total_messages": len(state.context) if state.context else 0, + **token_usage, + }, + thread_id=config.get("thread_id"), + run_id=config.get("run_id"), + ) diff --git a/agentflow/core/graph/utils/utils.py b/agentflow/core/graph/utils/utils.py index 6060310..b808471 100644 --- a/agentflow/core/graph/utils/utils.py +++ b/agentflow/core/graph/utils/utils.py @@ -50,6 +50,7 @@ async def parse_response( state: AgentState, messages: list[Message], response_granularity: ResponseGranularity = ResponseGranularity.LOW, + token_usage: dict[str, int] | None = None, ) -> dict[str, Any]: """Parse and format execution response based on specified granularity level. @@ -83,19 +84,20 @@ async def parse_response( match response_granularity: case ResponseGranularity.FULL: # Return full state and messages - return {"state": state, "messages": messages} + return {"state": state, "messages": messages, "token_usage": token_usage} case ResponseGranularity.PARTIAL: # Return state and summary of messages return { "context": state.context, "summary": state.context_summary, - "message": messages, + "messages": messages, + "token_usage": token_usage, } case ResponseGranularity.LOW: # Return all messages from state context - return {"messages": messages} + return {"messages": messages, "token_usage": token_usage} - return {"messages": messages} + return {"messages": messages, "token_usage": token_usage} # Utility to update only provided fields in state @@ -577,3 +579,60 @@ async def sync_data( await checkpointer.aput_messages(config, messages) return is_context_trimmed + + +def calculate_token_usage(messages: list[Message]) -> dict[str, int]: + """Calculate total token usage from all messages in the state. + + Aggregates token usage across all messages in the state's context, + including input tokens (prompt_tokens), output tokens (completion_tokens), + and reasoning tokens. + + Args: + messages: The list of messages containing token usage information. + + Returns: + Dictionary containing: + - total_input_tokens: Total prompt/input tokens used + - total_output_tokens: Total completion/output tokens used + - total_reasoning_tokens: Total reasoning tokens used + - total_tokens: Sum of input and output tokens + + Example: + ```python + usage = calculate_token_usage(state) + # Returns: { + # "total_input_tokens": 1500, + # "total_output_tokens": 800, + # "total_reasoning_tokens": 200, + # "total_tokens": 2300 + # } + ``` + """ + total_input_tokens = 0 + total_output_tokens = 0 + total_reasoning_tokens = 0 + + if not messages: + return { + "total_input_tokens": total_input_tokens, + "total_output_tokens": total_output_tokens, + "total_reasoning_tokens": total_reasoning_tokens, + "total_tokens": total_input_tokens + total_output_tokens, + } + + for message in messages: + if message.usages: + total_input_tokens += message.usages.prompt_tokens + total_output_tokens += message.usages.completion_tokens + total_reasoning_tokens += message.usages.reasoning_tokens + + # Note: total_tokens is input + output only (reasoning tracked separately) + total_tokens = total_input_tokens + total_output_tokens + + return { + "total_input_tokens": total_input_tokens, + "total_output_tokens": total_output_tokens, + "total_reasoning_tokens": total_reasoning_tokens, + "total_tokens": total_tokens, + } diff --git a/agentflow/runtime/protocols/__init__.py b/agentflow/runtime/protocols/__init__.py index c7914bb..11fec79 100644 --- a/agentflow/runtime/protocols/__init__.py +++ b/agentflow/runtime/protocols/__init__.py @@ -4,23 +4,22 @@ ``agentflow.runtime.protocols.a2a``. """ -from . import a2a -from .a2a import ( - AgentFlowExecutor, - build_a2a_app, - create_a2a_client_node, - create_a2a_server, - delegate_to_a2a_agent, - make_agent_card, -) +# from . import a2a +# from .a2a import ( +# AgentFlowExecutor, +# build_a2a_app, +# create_a2a_client_node, +# create_a2a_server, +# delegate_to_a2a_agent, +# make_agent_card, +# ) - -__all__ = [ - "AgentFlowExecutor", - "a2a", - "build_a2a_app", - "create_a2a_client_node", - "create_a2a_server", - "delegate_to_a2a_agent", - "make_agent_card", -] +# __all__ = [ +# "AgentFlowExecutor", +# "a2a", +# "build_a2a_app", +# "create_a2a_client_node", +# "create_a2a_server", +# "delegate_to_a2a_agent", +# "make_agent_card", +# ] diff --git a/agentflow/runtime/protocols/a2a/README.md b/agentflow/runtime/protocols/a2a/README.md deleted file mode 100644 index e5db421..0000000 --- a/agentflow/runtime/protocols/a2a/README.md +++ /dev/null @@ -1,445 +0,0 @@ -# agentflow.a2a_integration - -> **Official bridge between agentflow and the [A2A (Agent-to-Agent) protocol](https://github.com/google/A2A) via the `a2a-sdk`.** - -This package lets you **expose** any agentflow `CompiledGraph` as an A2A-compliant agent, and **call** remote A2A agents from within agentflow graphs — all with minimal boilerplate. - ---- - -## Table of Contents - -- [Installation](#installation) -- [Architecture Overview](#architecture-overview) -- [Module Reference](#module-reference) - - [executor.py — AgentFlowExecutor](#executorpy--agentflowexecutor) - - [server.py — Server Helpers](#serverpy--server-helpers) - - [client.py — Client Helpers](#clientpy--client-helpers) -- [Quick Start](#quick-start) - - [Serving a Graph as an A2A Agent](#serving-a-graph-as-an-a2a-agent) - - [Calling a Remote A2A Agent](#calling-a-remote-a2a-agent) - - [Using a Remote Agent as a Graph Node](#using-a-remote-agent-as-a-graph-node) -- [Conversation Memory (context_id)](#conversation-memory-context_id) -- [Streaming](#streaming) -- [Custom Executors](#custom-executors) -- [Examples](#examples) -- [API Summary](#api-summary) - ---- - -## Installation - -The A2A integration requires the `a2a-sdk` extra: - -```bash -pip install agentflow[a2a_sdk] -``` - -This installs `a2a-sdk`, `httpx`, `uvicorn`, and all transitive dependencies. - ---- - -## Architecture Overview - -``` -┌───────────────────────────────────────────────────────────────┐ -│ A2A Client │ -│ (any A2A-compliant client, browser, or another agent) │ -└──────────────┬────────────────────────────────────────────────┘ - │ JSON-RPC / SSE over HTTP - ▼ -┌──────────────────────────────────────────────────────────────┐ -│ a2a-sdk layer (transport, JSON-RPC, task lifecycle, SSE) │ -│ ┌────────────────────────────────────────────────────────┐ │ -│ │ DefaultRequestHandler + InMemoryTaskStore │ │ -│ └────────────────────┬───────────────────────────────────┘ │ -│ │ calls execute() │ -│ ┌────────────────────▼───────────────────────────────────┐ │ -│ │ AgentFlowExecutor (executor.py) │ │ -│ │ ─ extracts user text from A2A message parts │ │ -│ │ ─ resolves thread_id from context_id │ │ -│ │ ─ runs CompiledGraph.ainvoke() or .astream() │ │ -│ │ ─ pushes results back via TaskUpdater │ │ -│ └────────────────────┬───────────────────────────────────┘ │ -└───────────────────────┼──────────────────────────────────────┘ - │ - ┌────────▼────────┐ - │ CompiledGraph │ ← your agentflow graph - │ (any topology) │ - └─────────────────┘ -``` - -**Key insight:** The `a2a-sdk` owns all transport (HTTP, JSON-RPC, SSE, task state machines). Agentflow owns all agent logic (LLM calls, tool execution, state management, checkpointing). `AgentFlowExecutor` is the sole bridge connecting the two. - ---- - -## Module Reference - -### `executor.py` — AgentFlowExecutor - -The core bridge class. Implements the `a2a-sdk`'s `AgentExecutor` interface. - -```python -class AgentFlowExecutor(AgentExecutor): - def __init__( - self, - compiled_graph: CompiledGraph, - config: dict[str, Any] | None = None, - streaming: bool = False, - ) -> None: ... -``` - -| Parameter | Type | Description | -|-----------|------|-------------| -| `compiled_graph` | `CompiledGraph` | A fully compiled agentflow graph | -| `config` | `dict` | Base config forwarded to `ainvoke`/`astream` (e.g. `recursion_limit`) | -| `streaming` | `bool` | When `True`, uses `astream` + sends `TaskState.working` progress events | - -**What it does on each A2A request:** - -1. Creates a `TaskUpdater` and marks the task as `submitted` → `working` -2. Extracts user text from the A2A message parts via `context.get_user_input()` -3. Wraps the text as an agentflow `Message` list -4. Resolves `thread_id` from `context.context_id` (falling back to `context.task_id`) for checkpointer-based conversation memory -5. Calls `graph.ainvoke()` (blocking) or `graph.astream()` (streaming) -6. Extracts the last assistant message from the result state -7. Pushes it back as an A2A artifact via `updater.add_artifact()` and marks `completed` - -**Text extraction** walks backwards through `state.context` looking for the last `role="assistant"` message — this works regardless of how many nodes the graph has. - -**Error handling** catches all exceptions and pushes a `TaskState.failed` status with the error message. - ---- - -### `server.py` — Server Helpers - -Three convenience functions to expose a graph as an A2A HTTP server: - -#### `make_agent_card()` - -Builds an `AgentCard` (the A2A discovery descriptor served at `/.well-known/agent-card.json`). - -```python -def make_agent_card( - name: str, - description: str, - url: str, - *, - skills: list[AgentSkill] | None = None, - streaming: bool = False, - version: str = "1.0.0", -) -> AgentCard: ... -``` - -If no `skills` are provided, a default `"run_graph"` skill is created automatically. - -#### `build_a2a_app()` - -Returns a **Starlette ASGI app** — useful when you need to compose the A2A endpoint with other routes (e.g. mount inside FastAPI) or run it in tests. - -```python -def build_a2a_app( - compiled_graph: CompiledGraph, - agent_card: AgentCard, - *, - streaming: bool = False, - executor_config: dict[str, Any] | None = None, -) -> Starlette: ... -``` - -#### `create_a2a_server()` - -**One-call blocking server** — builds the app and starts `uvicorn`. Ideal for standalone agents. - -```python -def create_a2a_server( - compiled_graph: CompiledGraph, - agent_card: AgentCard, - *, - host: str = "0.0.0.0", - port: int = 9999, - streaming: bool = False, - executor_config: dict[str, Any] | None = None, -) -> None: ... -``` - ---- - -### `client.py` — Client Helpers - -Utilities for calling **remote** A2A agents from within agentflow graphs. - -#### `delegate_to_a2a_agent()` - -Async one-shot helper — send text, get text back. - -```python -async def delegate_to_a2a_agent( - url: str, - text: str, - *, - timeout: float = 30.0, -) -> str: ... -``` - -Sends a single `TextPart` message to the remote agent and returns the response text. Raises `RuntimeError` if the agent returns an error or no text content. - -#### `create_a2a_client_node()` - -Factory that returns an **agentflow graph node function** wrapping a remote A2A agent. The returned callable has the standard node signature `(state, config) -> list[Message]`. - -```python -def create_a2a_client_node( - url: str, - *, - timeout: float = 30.0, - response_role: str = "assistant", -) -> Callable: ... -``` - -**Usage in a graph:** - -```python -from agentflow.a2a_integration import create_a2a_client_node - -graph.add_node("remote_agent", create_a2a_client_node("http://localhost:9999")) -graph.add_edge("some_node", "remote_agent") -graph.add_edge("remote_agent", END) -``` - -The node reads the last message from `state.context`, forwards its text to the remote A2A agent, and returns the response as a new `Message`. - ---- - -## Quick Start - -### Serving a Graph as an A2A Agent - -```python -from agentflow.graph import StateGraph -from agentflow.state import AgentState -from agentflow.utils.constants import END -from agentflow.a2a_integration import ( - create_a2a_server, - make_agent_card, -) - -# 1. Build your agentflow graph -async def my_node(state: AgentState, config: dict): - from agentflow.state import Message - user_text = state.context[-1].text() if state.context else "" - return [Message.text_message(f"You said: {user_text}", role="assistant")] - -graph = StateGraph[AgentState](AgentState()) -graph.add_node("main", my_node) -graph.set_entry_point("main") -graph.add_edge("main", END) -compiled = graph.compile() - -# 2. Create the agent card -card = make_agent_card( - name="EchoAgent", - description="Echoes back whatever you say", - url="http://localhost:9999", -) - -# 3. Start the A2A server (blocking) -create_a2a_server(compiled, card, port=9999) -``` - -The agent is now discoverable at `http://localhost:9999/.well-known/agent-card.json` and accepts JSON-RPC requests at `http://localhost:9999/`. - -### Calling a Remote A2A Agent - -```python -from agentflow.a2a_integration import delegate_to_a2a_agent - -response = await delegate_to_a2a_agent( - "http://localhost:9999", - "Hello, agent!", -) -print(response) # "You said: Hello, agent!" -``` - -### Using a Remote Agent as a Graph Node - -```python -from agentflow.a2a_integration import create_a2a_client_node -from agentflow.graph import StateGraph -from agentflow.state import AgentState -from agentflow.utils.constants import END - -graph = StateGraph[AgentState](AgentState()) -graph.add_node("user_input", some_input_node) -graph.add_node("remote", create_a2a_client_node("http://localhost:9999")) -graph.set_entry_point("user_input") -graph.add_edge("user_input", "remote") -graph.add_edge("remote", END) - -compiled = graph.compile() -result = await compiled.ainvoke({"messages": [...]}) -``` - ---- - -## Conversation Memory (context_id) - -The A2A protocol has two key identifiers: - -| Field | Purpose | -|-------|---------| -| `task_id` | Unique per message/request — changes every call | -| `context_id` | Stable per conversation session — stays the same across turns | - -`AgentFlowExecutor` uses `context_id` as the agentflow checkpointer's `thread_id`. This means: - -- **Same `context_id`** → same checkpointer thread → conversation history is restored -- **Different `context_id`** → fresh thread → new conversation -- **No `context_id`** → falls back to `task_id` → one-shot (no memory) - -``` -Turn 1: context_id="abc" → thread_id="abc" → checkpointer saves state -Turn 2: context_id="abc" → thread_id="abc" → checkpointer restores state ✓ -Turn 3: context_id="xyz" → thread_id="xyz" → fresh conversation -``` - -The client is responsible for sending a consistent `context_id` across turns. The `a2a-sdk` `ClientFactory` handles this automatically when configured. - -### Sub-agent context_id isolation - -When an agent delegates to another agent (e.g. PlannerAgent → CurrencyAgent), **do not forward the caller's `context_id` directly**. Both agents would share the same checkpointer thread, causing their conversation histories to collide. - -Instead, derive a namespaced `context_id` for each sub-agent: - -```python -# Inside a delegation tool -planner_ctx_id = state.a2a_context_id if state else "" -currency_ctx_id = f"{planner_ctx_id}:currency" if planner_ctx_id else "" -result = await _send_to_currency_agent(query, context_id=currency_ctx_id) -``` - -This ensures: -- PlannerAgent checkpoints under `context_id` (e.g. `"abc"`) -- CurrencyAgent checkpoints under `"abc:currency"` -- Both maintain independent, stable conversation memory across turns -- The currency result is still returned to the planner's state as a tool message - ---- - -## Streaming - -When `streaming=True`, the executor uses `CompiledGraph.astream()` instead of `ainvoke()`: - -```python -# Server side -executor = AgentFlowExecutor(compiled, streaming=True) - -# Or via the server helper -create_a2a_server(compiled, card, streaming=True) -``` - -**What happens during streaming:** - -1. The graph yields `StreamChunk` objects as each node completes -2. For each chunk with a message, the executor pushes a `TaskState.working` status update via SSE -3. After the stream completes, the final text is emitted as an artifact with `TaskState.completed` - -The client observes real-time progress via Server-Sent Events (SSE). - -**Note:** This is A2A-level streaming — one SSE `TaskState.working` event is sent per `StreamEvent.MESSAGE` chunk emitted by the graph (i.e. whenever a node produces a message). Token-level LLM streaming is a separate concern from the A2A transport layer. - ---- - -## Custom Executors - -For advanced use cases (e.g. detecting `input_required`, merging custom state fields), subclass `AgentFlowExecutor`: - -```python -from agentflow.a2a_integration.executor import AgentFlowExecutor -from a2a.server.agent_execution.context import RequestContext -from a2a.server.events.event_queue import EventQueue -from a2a.server.tasks.task_updater import TaskUpdater -from a2a.types import TaskState, TextPart - -class MyCustomExecutor(AgentFlowExecutor): - async def execute(self, context: RequestContext, event_queue: EventQueue) -> None: - updater = TaskUpdater( - event_queue=event_queue, - task_id=context.task_id or "", - context_id=context.context_id or "", - ) - await updater.submit() - await updater.start_work() - - user_text = context.get_user_input() if context.message else "" - messages = [AFMessage.text_message(user_text, role="user")] - - run_config = { - "thread_id": context.context_id or context.task_id or "", - } - - result = await self.graph.ainvoke( - { - "messages": messages, - # Merge custom fields into your AgentState subclass - "state": {"my_custom_field": "some_value"}, - }, - config=run_config, - ) - - response_text = self._extract_response_text(result) - - # Custom logic: detect if agent needs user input - if self._needs_user_input(result): - msg = updater.new_agent_message(parts=[TextPart(text=response_text)]) - await updater.update_status(TaskState.input_required, message=msg) - else: - await updater.add_artifact([TextPart(text=response_text)]) - await updater.complete() -``` - -**Common customisation patterns:** - -| Pattern | How | -|---------|-----| -| **`input_required` relay** | Inspect `result["state"]` for tool/status markers, set `TaskState.input_required` | -| **Custom state fields** | Subclass `AgentState` and pass `"state": {field: value}` in the `ainvoke` input dict to merge extra fields (e.g. `PlannerState.a2a_context_id`) into the graph state before the first node runs | -| **Multi-agent delegation** | Use `delegate_to_a2a_agent()` or a custom tool inside a `ToolNode`; derive a namespaced `context_id` per sub-agent (e.g. `f"{ctx_id}:currency"`) to keep each agent's checkpointer thread isolated | -| **Task metadata** | Access `context.task_id`, `context.context_id`, `context.message` for routing decisions | - ---- - -## Examples - -Working examples are in `examples/a2a_sdk/`: - -| Example | Description | Key Concepts | -|---------|-------------|--------------| -| **`currency/`** | Single ReAct agent with currency conversion tool | `AgentFlowExecutor`, `make_agent_card`, `create_a2a_server` | -| **`pattern1_human_agent/`** | Human-in-the-loop with `input_required` | Custom executor, `TaskState.input_required`, conversation memory | -| **`pattern2_orchestrator/`** | Central orchestrator routing to specialist agents | Multi-agent coordination, `delegate_to_a2a_agent` | -| **`pattern3_smart_client/`** | Client-side routing to multiple agents | Client-driven orchestration, per-agent `context_id` | -| **`pattern4_planner_delegates/`** | Planner ReAct graph delegates currency queries to CurrencyAgent via a `ToolNode` tool; detects `input_required` and relays it across agent boundaries | Custom `PlannerState` carrying `a2a_context_id`, namespaced sub-agent `context_id` (`f"{ctx}:currency"`), `_tool_requested_input` inspection of state, `TaskState.input_required` relay | - ---- - -## API Summary - -### Public Exports (`from agentflow.a2a_integration import ...`) - -| Name | Type | Description | -|------|------|-------------| -| `AgentFlowExecutor` | Class | Bridges `CompiledGraph` into A2A's `AgentExecutor` interface | -| `make_agent_card` | Function | Builds an `AgentCard` with sensible defaults | -| `build_a2a_app` | Function | Returns a Starlette ASGI app (composable) | -| `create_a2a_server` | Function | One-call blocking server (uvicorn) | -| `delegate_to_a2a_agent` | Function | Async one-shot: send text to remote agent, get text back | -| `create_a2a_client_node` | Function | Factory returning a graph node that wraps a remote A2A agent | - -### Dependencies - -| Package | Purpose | -|---------|---------| -| `a2a-sdk` | A2A protocol implementation (transport, JSON-RPC, SSE, task lifecycle) | -| `httpx` | Async HTTP client for `delegate_to_a2a_agent` | -| `uvicorn` | ASGI server for `create_a2a_server` | -| `starlette` | ASGI framework (via `a2a-sdk`) | diff --git a/agentflow/runtime/protocols/a2a/__init__.py b/agentflow/runtime/protocols/a2a/__init__.py index 22697c2..33c2b27 100644 --- a/agentflow/runtime/protocols/a2a/__init__.py +++ b/agentflow/runtime/protocols/a2a/__init__.py @@ -1,36 +1,36 @@ -"""Optional A2A protocol bridge for Agentflow. +# """Optional A2A protocol bridge for Agentflow. -This package exposes any agentflow ``CompiledGraph`` as a standard A2A -agent using the official ``a2a-sdk`` package, and also provides client -helpers to call remote A2A agents from within a graph. +# This package exposes any agentflow ``CompiledGraph`` as a standard A2A +# agent using the official ``a2a-sdk`` package, and also provides client +# helpers to call remote A2A agents from within a graph. -Install the extra: +# Install the extra: - pip install 10xscale-agentflow[a2a_sdk] +# pip install 10xscale-agentflow[a2a_sdk] -Quick start - server: +# Quick start - server: - from agentflow.runtime.protocols.a2a import ( - AgentFlowExecutor, - create_a2a_server, - make_agent_card, - ) +# from agentflow.runtime.protocols.a2a import ( +# AgentFlowExecutor, +# create_a2a_server, +# make_agent_card, +# ) -Quick start - client: +# Quick start - client: - from agentflow.runtime.protocols.a2a import delegate_to_a2a_agent -""" +# from agentflow.runtime.protocols.a2a import delegate_to_a2a_agent +# """ -from .client import create_a2a_client_node, delegate_to_a2a_agent -from .executor import AgentFlowExecutor -from .server import build_a2a_app, create_a2a_server, make_agent_card +# from .client import create_a2a_client_node, delegate_to_a2a_agent +# from .executor import AgentFlowExecutor +# from .server import build_a2a_app, create_a2a_server, make_agent_card -__all__ = [ - "AgentFlowExecutor", - "build_a2a_app", - "create_a2a_client_node", - "create_a2a_server", - "delegate_to_a2a_agent", - "make_agent_card", -] +# __all__ = [ +# "AgentFlowExecutor", +# "build_a2a_app", +# "create_a2a_client_node", +# "create_a2a_server", +# "delegate_to_a2a_agent", +# "make_agent_card", +# ] diff --git a/agentflow/runtime/protocols/a2a/_optional.py b/agentflow/runtime/protocols/a2a/_optional.py index b624117..4012fc6 100644 --- a/agentflow/runtime/protocols/a2a/_optional.py +++ b/agentflow/runtime/protocols/a2a/_optional.py @@ -1,36 +1,36 @@ -"""Helpers for loading A2A optional dependencies.""" +# """Helpers for loading A2A optional dependencies.""" -from __future__ import annotations +# from __future__ import annotations -from importlib import import_module -from types import ModuleType -from typing import Any +# from importlib import import_module +# from types import ModuleType +# from typing import Any -A2A_EXTRA_INSTALL_HINT = ( - "Install it with 'pip install 10xscale-agentflow[a2a_sdk]' " "or 'pip install a2a-sdk'." -) +# A2A_EXTRA_INSTALL_HINT = ( +# "Install it with 'pip install 10xscale-agentflow[a2a_sdk]' " "or 'pip install a2a-sdk'." +# ) -def missing_a2a_sdk_error(feature: str, exc: BaseException) -> RuntimeError: - """Return a consistent error for A2A helpers when a2a-sdk is absent.""" - return RuntimeError( - f"{feature} requires the optional 'a2a-sdk' package. {A2A_EXTRA_INSTALL_HINT}" - ) +# def missing_a2a_sdk_error(feature: str, exc: BaseException) -> RuntimeError: +# """Return a consistent error for A2A helpers when a2a-sdk is absent.""" +# return RuntimeError( +# f"{feature} requires the optional 'a2a-sdk' package. {A2A_EXTRA_INSTALL_HINT}" +# ) -def import_a2a_module(module_name: str, feature: str) -> ModuleType: - """Import an A2A SDK module with a helpful optional-dependency error.""" - try: - return import_module(module_name) - except Exception as exc: - raise missing_a2a_sdk_error(feature, exc) from exc +# def import_a2a_module(module_name: str, feature: str) -> ModuleType: +# """Import an A2A SDK module with a helpful optional-dependency error.""" +# try: +# return import_module(module_name) +# except Exception as exc: +# raise missing_a2a_sdk_error(feature, exc) from exc -def get_a2a_attr(module_name: str, attr_name: str, feature: str) -> Any: - """Get an attribute from an A2A SDK module with a consistent error.""" - module = import_a2a_module(module_name, feature) - try: - return getattr(module, attr_name) - except AttributeError as exc: - raise missing_a2a_sdk_error(feature, exc) from exc +# def get_a2a_attr(module_name: str, attr_name: str, feature: str) -> Any: +# """Get an attribute from an A2A SDK module with a consistent error.""" +# module = import_a2a_module(module_name, feature) +# try: +# return getattr(module, attr_name) +# except AttributeError as exc: +# raise missing_a2a_sdk_error(feature, exc) from exc diff --git a/agentflow/runtime/protocols/a2a/client.py b/agentflow/runtime/protocols/a2a/client.py index 43daf86..05a0e59 100644 --- a/agentflow/runtime/protocols/a2a/client.py +++ b/agentflow/runtime/protocols/a2a/client.py @@ -1,222 +1,222 @@ -""" -A2A client helpers for agentflow. +# """ +# A2A client helpers for agentflow. -Provides utilities to call any remote A2A-compliant agent from within -an agentflow graph. +# Provides utilities to call any remote A2A-compliant agent from within +# an agentflow graph. -Functions: - delegate_to_a2a_agent — async one-shot: send text, get text back. - create_a2a_client_node — factory returning a graph-compatible node - function that delegates to a remote A2A - agent. -""" +# Functions: +# delegate_to_a2a_agent — async one-shot: send text, get text back. +# create_a2a_client_node — factory returning a graph-compatible node +# function that delegates to a remote A2A +# agent. +# """ -from __future__ import annotations +# from __future__ import annotations -import logging -import uuid -from importlib import import_module -from typing import Any - -from agentflow.core.state.agent_state import AgentState -from agentflow.core.state.message import Message as AFMessage +# import logging +# import uuid +# from importlib import import_module +# from typing import Any + +# from agentflow.core.state.agent_state import AgentState +# from agentflow.core.state.message import Message as AFMessage -from ._optional import A2A_EXTRA_INSTALL_HINT, get_a2a_attr, import_a2a_module - - -logger = logging.getLogger("agentflow.a2a") - - -def _import_client_dependencies(): - """Load client-only optional dependencies when an A2A call is made.""" - feature = "A2A client helpers" - try: - httpx = import_module("httpx") - except Exception as exc: - raise RuntimeError( - f"{feature} requires the optional 'httpx' package. {A2A_EXTRA_INSTALL_HINT}" - ) from exc +# from ._optional import A2A_EXTRA_INSTALL_HINT, get_a2a_attr, import_a2a_module + + +# logger = logging.getLogger("agentflow.a2a") + + +# def _import_client_dependencies(): +# """Load client-only optional dependencies when an A2A call is made.""" +# feature = "A2A client helpers" +# try: +# httpx = import_module("httpx") +# except Exception as exc: +# raise RuntimeError( +# f"{feature} requires the optional 'httpx' package. {A2A_EXTRA_INSTALL_HINT}" +# ) from exc - a2a_client = get_a2a_attr("a2a.client", "A2AClient", feature) - a2a_types = import_a2a_module("a2a.types", feature) - return httpx, a2a_client, a2a_types +# a2a_client = get_a2a_attr("a2a.client", "A2AClient", feature) +# a2a_types = import_a2a_module("a2a.types", feature) +# return httpx, a2a_client, a2a_types -# ---------------------------------------------------------------------- # -# Low-level helper # -# ---------------------------------------------------------------------- # - - -async def delegate_to_a2a_agent( - url: str, - text: str, - *, - context_id: str | None = None, - timeout: float = 30.0, -) -> str: - """Call a remote A2A agent and return its text response. +# # ---------------------------------------------------------------------- # +# # Low-level helper # +# # ---------------------------------------------------------------------- # + + +# async def delegate_to_a2a_agent( +# url: str, +# text: str, +# *, +# context_id: str | None = None, +# timeout: float = 30.0, +# ) -> str: +# """Call a remote A2A agent and return its text response. - This uses the (deprecated but stable) ``A2AClient`` from the a2a-sdk - which provides the simplest request/response interface. +# This uses the (deprecated but stable) ``A2AClient`` from the a2a-sdk +# which provides the simplest request/response interface. - Args: - url: Base URL of the remote agent (e.g. ``http://localhost:9999``). - text: The user message to send. - timeout: HTTP request timeout in seconds. - - Returns: - The text content of the agent's response. - - Raises: - RuntimeError: If the agent returns an error or no text parts. - """ - httpx, a2a_client, a2a_types = _import_client_dependencies() - - async with httpx.AsyncClient(timeout=timeout) as http: - client = a2a_client(httpx_client=http, url=url) - - request = a2a_types.SendMessageRequest( - id=str(uuid.uuid4()), - params=a2a_types.MessageSendParams( - message=a2a_types.Message( - role=a2a_types.Role.user, - message_id=str(uuid.uuid4()), - context_id=context_id, - parts=[a2a_types.TextPart(text=text)], - ), - ), - ) - - response = await client.send_message(request) - - # response.root is either SendMessageSuccessResponse or JSONRPCErrorResponse - result = response.root - if hasattr(result, "error"): - raise RuntimeError(f"A2A agent returned error: {result.error}") - - # result.result is Task | Message - payload = result.result - - # Extract text from the response - return _extract_text(payload) - - -def _extract_text(payload: Any) -> str: - """Pull text from a Task or Message returned by the A2A SDK. - - The SDK wraps parts in ``Part(root=TextPart(...))`` — a discriminated - union. We check both ``part.text`` (direct TextPart) and - ``part.root.text`` (wrapped Part) to be resilient. - """ - parts: list[Any] = [] - - if hasattr(payload, "parts"): - # It's an A2A Message - parts = payload.parts or [] - elif hasattr(payload, "artifacts") and payload.artifacts: - # It's a Task — text lives in artifact parts - for artifact in payload.artifacts: - parts.extend(artifact.parts or []) - elif hasattr(payload, "status") and payload.status and payload.status.message: - # Fallback: check status message - parts = payload.status.message.parts or [] - - text_parts: list[str] = [] - for p in parts: - # Direct TextPart (has .text) - if hasattr(p, "text") and isinstance(p.text, str): - text_parts.append(p.text) - # Wrapped Part(root=TextPart(...)) - elif hasattr(p, "root") and hasattr(p.root, "text") and isinstance(p.root.text, str): - text_parts.append(p.root.text) - - if text_parts: - return "\n".join(text_parts) - - raise RuntimeError("A2A agent response contained no text parts") - - -# ---------------------------------------------------------------------- # -# Graph node factory # -# ---------------------------------------------------------------------- # - - -def create_a2a_client_node( - url: str, - *, - timeout: float = 30.0, - response_role: str = "assistant", -): - """Return an async callable that can be used as an agentflow graph node. - - The node reads the last message from the state, forwards its text to - the remote A2A agent at *url*, and returns the response as a new - ``Message``. - - Usage:: - - graph.add_node("remote_agent", create_a2a_client_node("http://localhost:9999")) - graph.add_edge("some_node", "remote_agent") - graph.add_edge("remote_agent", END) - - Args: - url: Base URL of the remote A2A agent. - timeout: HTTP request timeout. - response_role: Role to assign to the response message - (default ``"assistant"``). - - Returns: - An async function with signature - ``(state: AgentState, config: dict) -> list[AFMessage]`` - """ - - async def _a2a_node(state: AgentState, config: dict) -> list[AFMessage]: - # Get text from the last message in the conversation - if not state.context: - return [ - AFMessage.text_message( - "No input provided.", - role=response_role, - ), - ] - - user_text = state.context[-1].text() - if not user_text: - return [ - AFMessage.text_message( - "Empty input.", - role=response_role, - ), - ] - - # Reuse the parent graph's thread_id as context_id so the remote - # A2A agent stays in the same session as the whole workflow. - # The server uses context_id as its own thread_id for its checkpointer, - # so it maintains full conversation history server-side across turns. - context_id = config.get("thread_id") - - try: - response = await delegate_to_a2a_agent( - url, user_text, context_id=context_id, timeout=timeout - ) - except Exception as exc: - logger.exception("A2A client node failed for url=%s", url) - return [ - AFMessage.text_message( - f"A2A call failed: {exc!s}", - role=response_role, - ), - ] - - return [ - AFMessage.text_message( - response, - role=response_role, - ), - ] - - # Give the function a useful name for debugging / graph visualization - _a2a_node.__name__ = f"a2a_client_node({url})" - _a2a_node.__qualname__ = _a2a_node.__name__ - - return _a2a_node +# Args: +# url: Base URL of the remote agent (e.g. ``http://localhost:9999``). +# text: The user message to send. +# timeout: HTTP request timeout in seconds. + +# Returns: +# The text content of the agent's response. + +# Raises: +# RuntimeError: If the agent returns an error or no text parts. +# """ +# httpx, a2a_client, a2a_types = _import_client_dependencies() + +# async with httpx.AsyncClient(timeout=timeout) as http: +# client = a2a_client(httpx_client=http, url=url) + +# request = a2a_types.SendMessageRequest( +# id=str(uuid.uuid4()), +# params=a2a_types.MessageSendParams( +# message=a2a_types.Message( +# role=a2a_types.Role.user, +# message_id=str(uuid.uuid4()), +# context_id=context_id, +# parts=[a2a_types.TextPart(text=text)], +# ), +# ), +# ) + +# response = await client.send_message(request) + +# # response.root is either SendMessageSuccessResponse or JSONRPCErrorResponse +# result = response.root +# if hasattr(result, "error"): +# raise RuntimeError(f"A2A agent returned error: {result.error}") + +# # result.result is Task | Message +# payload = result.result + +# # Extract text from the response +# return _extract_text(payload) + + +# def _extract_text(payload: Any) -> str: +# """Pull text from a Task or Message returned by the A2A SDK. + +# The SDK wraps parts in ``Part(root=TextPart(...))`` — a discriminated +# union. We check both ``part.text`` (direct TextPart) and +# ``part.root.text`` (wrapped Part) to be resilient. +# """ +# parts: list[Any] = [] + +# if hasattr(payload, "parts"): +# # It's an A2A Message +# parts = payload.parts or [] +# elif hasattr(payload, "artifacts") and payload.artifacts: +# # It's a Task — text lives in artifact parts +# for artifact in payload.artifacts: +# parts.extend(artifact.parts or []) +# elif hasattr(payload, "status") and payload.status and payload.status.message: +# # Fallback: check status message +# parts = payload.status.message.parts or [] + +# text_parts: list[str] = [] +# for p in parts: +# # Direct TextPart (has .text) +# if hasattr(p, "text") and isinstance(p.text, str): +# text_parts.append(p.text) +# # Wrapped Part(root=TextPart(...)) +# elif hasattr(p, "root") and hasattr(p.root, "text") and isinstance(p.root.text, str): +# text_parts.append(p.root.text) + +# if text_parts: +# return "\n".join(text_parts) + +# raise RuntimeError("A2A agent response contained no text parts") + + +# # ---------------------------------------------------------------------- # +# # Graph node factory # +# # ---------------------------------------------------------------------- # + + +# def create_a2a_client_node( +# url: str, +# *, +# timeout: float = 30.0, +# response_role: str = "assistant", +# ): +# """Return an async callable that can be used as an agentflow graph node. + +# The node reads the last message from the state, forwards its text to +# the remote A2A agent at *url*, and returns the response as a new +# ``Message``. + +# Usage:: + +# graph.add_node("remote_agent", create_a2a_client_node("http://localhost:9999")) +# graph.add_edge("some_node", "remote_agent") +# graph.add_edge("remote_agent", END) + +# Args: +# url: Base URL of the remote A2A agent. +# timeout: HTTP request timeout. +# response_role: Role to assign to the response message +# (default ``"assistant"``). + +# Returns: +# An async function with signature +# ``(state: AgentState, config: dict) -> list[AFMessage]`` +# """ + +# async def _a2a_node(state: AgentState, config: dict) -> list[AFMessage]: +# # Get text from the last message in the conversation +# if not state.context: +# return [ +# AFMessage.text_message( +# "No input provided.", +# role=response_role, +# ), +# ] + +# user_text = state.context[-1].text() +# if not user_text: +# return [ +# AFMessage.text_message( +# "Empty input.", +# role=response_role, +# ), +# ] + +# # Reuse the parent graph's thread_id as context_id so the remote +# # A2A agent stays in the same session as the whole workflow. +# # The server uses context_id as its own thread_id for its checkpointer, +# # so it maintains full conversation history server-side across turns. +# context_id = config.get("thread_id") + +# try: +# response = await delegate_to_a2a_agent( +# url, user_text, context_id=context_id, timeout=timeout +# ) +# except Exception as exc: +# logger.exception("A2A client node failed for url=%s", url) +# return [ +# AFMessage.text_message( +# f"A2A call failed: {exc!s}", +# role=response_role, +# ), +# ] + +# return [ +# AFMessage.text_message( +# response, +# role=response_role, +# ), +# ] + +# # Give the function a useful name for debugging / graph visualization +# _a2a_node.__name__ = f"a2a_client_node({url})" +# _a2a_node.__qualname__ = _a2a_node.__name__ + +# return _a2a_node diff --git a/agentflow/runtime/protocols/a2a/executor.py b/agentflow/runtime/protocols/a2a/executor.py index e138543..41f9b7d 100644 --- a/agentflow/runtime/protocols/a2a/executor.py +++ b/agentflow/runtime/protocols/a2a/executor.py @@ -1,224 +1,224 @@ -""" -AgentFlowExecutor — the sole bridge between agentflow and the a2a-sdk. +# """ +# AgentFlowExecutor — the sole bridge between agentflow and the a2a-sdk. -This module implements the ``AgentExecutor`` interface from the official -``a2a-sdk`` so that **any** agentflow ``CompiledGraph`` can be served as -a standard A2A agent. The SDK handles all HTTP, JSON-RPC, SSE, and task -lifecycle concerns. Agentflow handles all agent logic. - -Blocking path (default): - Uses ``CompiledGraph.ainvoke`` to run the graph to completion, then - emits a single ``TextPart`` artifact with the last assistant message. - -Streaming path (``streaming=True``): - Uses ``CompiledGraph.astream`` to yield incremental ``StreamChunk`` - objects. For each chunk that carries a message, the executor sends a - ``TaskState.working`` status update so the A2A client can observe - progress. After the stream finishes the final text is emitted as an - artifact. -""" - -from __future__ import annotations - -import logging -from typing import TYPE_CHECKING, Any - -from agentflow.core.state.message import Message as AFMessage -from agentflow.core.state.stream_chunks import StreamEvent -from agentflow.utils.constants import ResponseGranularity - -from ._optional import missing_a2a_sdk_error - - -if TYPE_CHECKING: - from agentflow.core.graph.compiled_graph import CompiledGraph - - -try: - from a2a.server.agent_execution import AgentExecutor as _AgentExecutor - from a2a.server.agent_execution.context import RequestContext - from a2a.server.events.event_queue import EventQueue - from a2a.server.tasks.task_updater import TaskUpdater - from a2a.types import TaskState, TextPart -except Exception as exc: - _A2A_IMPORT_ERROR: BaseException | None = exc - - class _AgentExecutor: - """Fallback base class used when a2a-sdk is not installed.""" - -else: - _A2A_IMPORT_ERROR = None - -AgentExecutor = _AgentExecutor - - -logger = logging.getLogger("agentflow.a2a") - - -class AgentFlowExecutor(AgentExecutor): - """Bridges a :class:`CompiledGraph` into the A2A execution model. - - This is the **only** glue code needed between agentflow and a2a-sdk. - The SDK owns the transport; agentflow owns the agent logic. - - Args: - compiled_graph: A fully compiled agentflow graph. - config: Optional base config dict forwarded to ``ainvoke`` / - ``astream`` (e.g. ``thread_id``, ``recursion_limit``). - streaming: When *True* the executor uses ``astream`` instead of - ``ainvoke`` and sends ``TaskState.working`` progress events. - """ - - def __init__( - self, - compiled_graph: CompiledGraph, - config: dict[str, Any] | None = None, - streaming: bool = False, - ) -> None: - if _A2A_IMPORT_ERROR is not None: - raise missing_a2a_sdk_error("AgentFlowExecutor", _A2A_IMPORT_ERROR) from ( - _A2A_IMPORT_ERROR - ) - - self.graph = compiled_graph - self._base_config = config or {} - self._streaming = streaming - - # ------------------------------------------------------------------ # - # A2A AgentExecutor interface # - # ------------------------------------------------------------------ # - - async def execute( - self, - context: RequestContext, - event_queue: EventQueue, - ) -> None: - """Run the agentflow graph for an incoming A2A request. - - 1. Extract user text from the A2A message parts. - 2. Run the graph (blocking or streaming). - 3. Push the result back as an A2A artifact. - """ - updater = TaskUpdater( - event_queue=event_queue, - task_id=context.task_id or "", - context_id=context.context_id or "", - ) - await updater.submit() - await updater.start_work() - - try: - # --- extract user text from A2A message parts ---------------- - user_text = context.get_user_input() if context.message else "" - - # build agentflow messages list - messages = [AFMessage.text_message(user_text, role="user")] - - # per-request config — use context_id as thread_id so that - # conversation history persists across A2A tasks within the - # same session. Falls back to task_id for one-shot callers. - run_config: dict[str, Any] = {**self._base_config} - if "thread_id" not in run_config: - run_config["thread_id"] = context.context_id or context.task_id or "" - - if self._streaming: - response_text = await self._execute_streaming(messages, run_config, updater) - else: - response_text = await self._execute_blocking(messages, run_config) - - # --- emit the final artifact --------------------------------- - await updater.add_artifact([TextPart(text=response_text)]) - await updater.complete() - - except Exception as exc: - logger.exception("AgentFlowExecutor: graph execution failed") - error_msg = updater.new_agent_message(parts=[TextPart(text=f"Error: {exc!s}")]) - await updater.failed(message=error_msg) - - async def cancel( - self, - context: RequestContext, - event_queue: EventQueue, - ) -> None: - """Cancel is not currently supported by agentflow graphs.""" - raise NotImplementedError("cancel not supported") - - # ------------------------------------------------------------------ # - # Internal helpers # - # ------------------------------------------------------------------ # - - async def _execute_blocking( - self, - messages: list[AFMessage], - config: dict[str, Any], - ) -> str: - """Run the graph via ``ainvoke`` and return the last assistant text.""" - result = await self.graph.ainvoke( - {"messages": messages}, - config=config, - response_granularity=ResponseGranularity.FULL, - ) - return self._extract_response_text(result) - - async def _execute_streaming( - self, - messages: list[AFMessage], - config: dict[str, Any], - updater: TaskUpdater, - ) -> str: - """Run the graph via ``astream``, sending progress updates per chunk.""" - last_text = "" - async for chunk in self.graph.astream( - {"messages": messages}, - config=config, - response_granularity=ResponseGranularity.FULL, - ): - if chunk.event == StreamEvent.MESSAGE and chunk.message is not None: - text = chunk.message.text() - if text: - last_text = text - # signal progress with the latest text - progress_msg = updater.new_agent_message(parts=[TextPart(text=text)]) - await updater.update_status(TaskState.working, message=progress_msg) - elif chunk.event == StreamEvent.STATE and chunk.state is not None: - # final state arrived — extract from it - assistant_text = self._extract_state_text(chunk.state) - if assistant_text: - last_text = assistant_text - - return last_text or "No response generated." - - # ------------------------------------------------------------------ # - # Text extraction # - # ------------------------------------------------------------------ # - - @staticmethod - def _extract_response_text(result: dict[str, Any]) -> str: - """Pull the last assistant message text from an ``ainvoke`` result. - - With ``ResponseGranularity.FULL`` the result dict contains - ``"state"`` (the complete ``AgentState``) as well as - ``"messages"`` (last-step messages). We prefer the full state - because it has all messages across every node. - """ - # Primary: full state context - full_state = result.get("state") - if full_state is not None: - for msg in reversed(full_state.context): - if msg.role == "assistant": - return msg.text() or "" - - # Fallback: messages list at LOW/PARTIAL granularity - for msg in reversed(result.get("messages", [])): - if msg.role == "assistant": - return msg.text() or "" - - return "No response generated." - - @staticmethod - def _extract_state_text(state: Any) -> str: - """Extract last assistant text from an ``AgentState``.""" - for msg in reversed(state.context): - if msg.role == "assistant": - return msg.text() or "" - return "" +# This module implements the ``AgentExecutor`` interface from the official +# ``a2a-sdk`` so that **any** agentflow ``CompiledGraph`` can be served as +# a standard A2A agent. The SDK handles all HTTP, JSON-RPC, SSE, and task +# lifecycle concerns. Agentflow handles all agent logic. + +# Blocking path (default): +# Uses ``CompiledGraph.ainvoke`` to run the graph to completion, then +# emits a single ``TextPart`` artifact with the last assistant message. + +# Streaming path (``streaming=True``): +# Uses ``CompiledGraph.astream`` to yield incremental ``StreamChunk`` +# objects. For each chunk that carries a message, the executor sends a +# ``TaskState.working`` status update so the A2A client can observe +# progress. After the stream finishes the final text is emitted as an +# artifact. +# """ + +# from __future__ import annotations + +# import logging +# from typing import TYPE_CHECKING, Any + +# from agentflow.core.state.message import Message as AFMessage +# from agentflow.core.state.stream_chunks import StreamEvent +# from agentflow.utils.constants import ResponseGranularity + +# from ._optional import missing_a2a_sdk_error + + +# if TYPE_CHECKING: +# from agentflow.core.graph.compiled_graph import CompiledGraph + + +# try: +# from a2a.server.agent_execution import AgentExecutor as _AgentExecutor +# from a2a.server.agent_execution.context import RequestContext +# from a2a.server.events.event_queue import EventQueue +# from a2a.server.tasks.task_updater import TaskUpdater +# from a2a.types import TaskState, TextPart +# except Exception as exc: +# _A2A_IMPORT_ERROR: BaseException | None = exc + +# class _AgentExecutor: +# """Fallback base class used when a2a-sdk is not installed.""" + +# else: +# _A2A_IMPORT_ERROR = None + +# AgentExecutor = _AgentExecutor + + +# logger = logging.getLogger("agentflow.a2a") + + +# class AgentFlowExecutor(AgentExecutor): +# """Bridges a :class:`CompiledGraph` into the A2A execution model. + +# This is the **only** glue code needed between agentflow and a2a-sdk. +# The SDK owns the transport; agentflow owns the agent logic. + +# Args: +# compiled_graph: A fully compiled agentflow graph. +# config: Optional base config dict forwarded to ``ainvoke`` / +# ``astream`` (e.g. ``thread_id``, ``recursion_limit``). +# streaming: When *True* the executor uses ``astream`` instead of +# ``ainvoke`` and sends ``TaskState.working`` progress events. +# """ + +# def __init__( +# self, +# compiled_graph: CompiledGraph, +# config: dict[str, Any] | None = None, +# streaming: bool = False, +# ) -> None: +# if _A2A_IMPORT_ERROR is not None: +# raise missing_a2a_sdk_error("AgentFlowExecutor", _A2A_IMPORT_ERROR) from ( +# _A2A_IMPORT_ERROR +# ) + +# self.graph = compiled_graph +# self._base_config = config or {} +# self._streaming = streaming + +# # ------------------------------------------------------------------ # +# # A2A AgentExecutor interface # +# # ------------------------------------------------------------------ # + +# async def execute( +# self, +# context: RequestContext, +# event_queue: EventQueue, +# ) -> None: +# """Run the agentflow graph for an incoming A2A request. + +# 1. Extract user text from the A2A message parts. +# 2. Run the graph (blocking or streaming). +# 3. Push the result back as an A2A artifact. +# """ +# updater = TaskUpdater( +# event_queue=event_queue, +# task_id=context.task_id or "", +# context_id=context.context_id or "", +# ) +# await updater.submit() +# await updater.start_work() + +# try: +# # --- extract user text from A2A message parts ---------------- +# user_text = context.get_user_input() if context.message else "" + +# # build agentflow messages list +# messages = [AFMessage.text_message(user_text, role="user")] + +# # per-request config — use context_id as thread_id so that +# # conversation history persists across A2A tasks within the +# # same session. Falls back to task_id for one-shot callers. +# run_config: dict[str, Any] = {**self._base_config} +# if "thread_id" not in run_config: +# run_config["thread_id"] = context.context_id or context.task_id or "" + +# if self._streaming: +# response_text = await self._execute_streaming(messages, run_config, updater) +# else: +# response_text = await self._execute_blocking(messages, run_config) + +# # --- emit the final artifact --------------------------------- +# await updater.add_artifact([TextPart(text=response_text)]) +# await updater.complete() + +# except Exception as exc: +# logger.exception("AgentFlowExecutor: graph execution failed") +# error_msg = updater.new_agent_message(parts=[TextPart(text=f"Error: {exc!s}")]) +# await updater.failed(message=error_msg) + +# async def cancel( +# self, +# context: RequestContext, +# event_queue: EventQueue, +# ) -> None: +# """Cancel is not currently supported by agentflow graphs.""" +# raise NotImplementedError("cancel not supported") + +# # ------------------------------------------------------------------ # +# # Internal helpers # +# # ------------------------------------------------------------------ # + +# async def _execute_blocking( +# self, +# messages: list[AFMessage], +# config: dict[str, Any], +# ) -> str: +# """Run the graph via ``ainvoke`` and return the last assistant text.""" +# result = await self.graph.ainvoke( +# {"messages": messages}, +# config=config, +# response_granularity=ResponseGranularity.FULL, +# ) +# return self._extract_response_text(result) + +# async def _execute_streaming( +# self, +# messages: list[AFMessage], +# config: dict[str, Any], +# updater: TaskUpdater, +# ) -> str: +# """Run the graph via ``astream``, sending progress updates per chunk.""" +# last_text = "" +# async for chunk in self.graph.astream( +# {"messages": messages}, +# config=config, +# response_granularity=ResponseGranularity.FULL, +# ): +# if chunk.event == StreamEvent.MESSAGE and chunk.message is not None: +# text = chunk.message.text() +# if text: +# last_text = text +# # signal progress with the latest text +# progress_msg = updater.new_agent_message(parts=[TextPart(text=text)]) +# await updater.update_status(TaskState.working, message=progress_msg) +# elif chunk.event == StreamEvent.STATE and chunk.state is not None: +# # final state arrived — extract from it +# assistant_text = self._extract_state_text(chunk.state) +# if assistant_text: +# last_text = assistant_text + +# return last_text or "No response generated." + +# # ------------------------------------------------------------------ # +# # Text extraction # +# # ------------------------------------------------------------------ # + +# @staticmethod +# def _extract_response_text(result: dict[str, Any]) -> str: +# """Pull the last assistant message text from an ``ainvoke`` result. + +# With ``ResponseGranularity.FULL`` the result dict contains +# ``"state"`` (the complete ``AgentState``) as well as +# ``"messages"`` (last-step messages). We prefer the full state +# because it has all messages across every node. +# """ +# # Primary: full state context +# full_state = result.get("state") +# if full_state is not None: +# for msg in reversed(full_state.context): +# if msg.role == "assistant": +# return msg.text() or "" + +# # Fallback: messages list at LOW/PARTIAL granularity +# for msg in reversed(result.get("messages", [])): +# if msg.role == "assistant": +# return msg.text() or "" + +# return "No response generated." + +# @staticmethod +# def _extract_state_text(state: Any) -> str: +# """Extract last assistant text from an ``AgentState``.""" +# for msg in reversed(state.context): +# if msg.role == "assistant": +# return msg.text() or "" +# return "" diff --git a/agentflow/runtime/protocols/a2a/server.py b/agentflow/runtime/protocols/a2a/server.py index 5836e32..477a33f 100644 --- a/agentflow/runtime/protocols/a2a/server.py +++ b/agentflow/runtime/protocols/a2a/server.py @@ -1,172 +1,172 @@ -""" -A2A server helpers for agentflow. - -Provides convenience functions to expose a :class:`CompiledGraph` as an -A2A-compliant HTTP endpoint using the official ``a2a-sdk``. - -Functions: - create_a2a_server — one-call to start a uvicorn server. - build_a2a_app — returns a Starlette ASGI app (composable). - make_agent_card — builds an ``AgentCard`` with sensible defaults. -""" - -from __future__ import annotations - -from typing import TYPE_CHECKING, Any - -from ._optional import A2A_EXTRA_INSTALL_HINT, import_a2a_module - - -if TYPE_CHECKING: - from a2a.types import AgentCard, AgentSkill - from starlette.applications import Starlette - - from agentflow.core.graph.compiled_graph import CompiledGraph - - -# ---------------------------------------------------------------------- # -# AgentCard helper # -# ---------------------------------------------------------------------- # - - -def make_agent_card( - name: str, - description: str, - url: str, - *, - skills: list[AgentSkill] | None = None, - streaming: bool = False, - version: str = "1.0.0", -) -> AgentCard: - """Build an :class:`AgentCard` with sensible defaults. - - If *skills* is ``None`` a single ``"run_graph"`` skill is created - automatically. - - Args: - name: Human-readable agent name. - description: Short description of what the agent does. - url: Public URL where the agent is reachable. - skills: Optional list of ``AgentSkill`` objects. - streaming: Whether the agent supports SSE streaming. - version: Semantic version string. - - Returns: - A fully populated ``AgentCard``. - """ - a2a_types = import_a2a_module("a2a.types", "make_agent_card") - - if skills is None: - skills = [ - a2a_types.AgentSkill( - id="run_graph", - name="Run Graph", - description="Execute the agentflow graph", - tags=["agentflow"], - ) - ] - - return a2a_types.AgentCard( - name=name, - description=description, - url=url, - version=version, - capabilities=a2a_types.AgentCapabilities(streaming=streaming), - default_input_modes=["text"], - default_output_modes=["text"], - skills=skills, - ) - - -# ---------------------------------------------------------------------- # -# ASGI app builder # -# ---------------------------------------------------------------------- # - - -def build_a2a_app( - compiled_graph: CompiledGraph, - agent_card: AgentCard, - *, - streaming: bool = False, - executor_config: dict[str, Any] | None = None, -) -> Starlette: - """Return a Starlette ASGI app that speaks the A2A protocol. - - Useful when you want to mount the app inside another ASGI framework - (e.g. FastAPI), run it with a custom server, or use it in tests. - - Args: - compiled_graph: A compiled agentflow graph. - agent_card: The ``AgentCard`` describing this agent. - streaming: Whether to use ``astream`` vs ``ainvoke`` in the - executor. - executor_config: Optional base config forwarded to the graph - (e.g. ``{"recursion_limit": 50}``). - - Returns: - A ``Starlette`` application ready to be served. - """ - a2a_apps = import_a2a_module("a2a.server.apps", "build_a2a_app") - request_handlers = import_a2a_module("a2a.server.request_handlers", "build_a2a_app") - tasks = import_a2a_module("a2a.server.tasks", "build_a2a_app") - - from .executor import AgentFlowExecutor - - executor = AgentFlowExecutor( - compiled_graph, - config=executor_config, - streaming=streaming, - ) - handler = request_handlers.DefaultRequestHandler( - agent_executor=executor, - task_store=tasks.InMemoryTaskStore(), - ) - a2a_app = a2a_apps.A2AStarletteApplication( - agent_card=agent_card, - http_handler=handler, - ) - return a2a_app.build() - - -# ---------------------------------------------------------------------- # -# One-call server # -# ---------------------------------------------------------------------- # - - -def create_a2a_server( - compiled_graph: CompiledGraph, - agent_card: AgentCard, - *, - host: str = "127.0.0.1", - port: int = 9999, - streaming: bool = False, - executor_config: dict[str, Any] | None = None, -) -> None: - """Build and run an A2A server exposing the given graph. - - This is a blocking call — it starts uvicorn and does not return until - the server is shut down. - - Args: - compiled_graph: A compiled agentflow graph. - agent_card: The ``AgentCard`` describing this agent. - host: Bind address. - port: Bind port. - streaming: Whether to use ``astream`` in the executor. - executor_config: Optional base config forwarded to the graph. - """ - try: - uvicorn = __import__("uvicorn") - except Exception as exc: - raise RuntimeError( - "create_a2a_server requires the optional 'uvicorn' package. " - f"{A2A_EXTRA_INSTALL_HINT}" - ) from exc - - app = build_a2a_app( - compiled_graph, - agent_card, - streaming=streaming, - executor_config=executor_config, - ) - uvicorn.run(app, host=host, port=port) +# """ +# A2A server helpers for agentflow. + +# Provides convenience functions to expose a :class:`CompiledGraph` as an +# A2A-compliant HTTP endpoint using the official ``a2a-sdk``. + +# Functions: +# create_a2a_server — one-call to start a uvicorn server. +# build_a2a_app — returns a Starlette ASGI app (composable). +# make_agent_card — builds an ``AgentCard`` with sensible defaults. +# """ + +# from __future__ import annotations + +# from typing import TYPE_CHECKING, Any + +# from ._optional import A2A_EXTRA_INSTALL_HINT, import_a2a_module + + +# if TYPE_CHECKING: +# from a2a.types import AgentCard, AgentSkill +# from starlette.applications import Starlette + +# from agentflow.core.graph.compiled_graph import CompiledGraph + + +# # ---------------------------------------------------------------------- # +# # AgentCard helper # +# # ---------------------------------------------------------------------- # + + +# def make_agent_card( +# name: str, +# description: str, +# url: str, +# *, +# skills: list[AgentSkill] | None = None, +# streaming: bool = False, +# version: str = "1.0.0", +# ) -> AgentCard: +# """Build an :class:`AgentCard` with sensible defaults. + +# If *skills* is ``None`` a single ``"run_graph"`` skill is created +# automatically. + +# Args: +# name: Human-readable agent name. +# description: Short description of what the agent does. +# url: Public URL where the agent is reachable. +# skills: Optional list of ``AgentSkill`` objects. +# streaming: Whether the agent supports SSE streaming. +# version: Semantic version string. + +# Returns: +# A fully populated ``AgentCard``. +# """ +# a2a_types = import_a2a_module("a2a.types", "make_agent_card") + +# if skills is None: +# skills = [ +# a2a_types.AgentSkill( +# id="run_graph", +# name="Run Graph", +# description="Execute the agentflow graph", +# tags=["agentflow"], +# ) +# ] + +# return a2a_types.AgentCard( +# name=name, +# description=description, +# url=url, +# version=version, +# capabilities=a2a_types.AgentCapabilities(streaming=streaming), +# default_input_modes=["text"], +# default_output_modes=["text"], +# skills=skills, +# ) + + +# # ---------------------------------------------------------------------- # +# # ASGI app builder # +# # ---------------------------------------------------------------------- # + + +# def build_a2a_app( +# compiled_graph: CompiledGraph, +# agent_card: AgentCard, +# *, +# streaming: bool = False, +# executor_config: dict[str, Any] | None = None, +# ) -> Starlette: +# """Return a Starlette ASGI app that speaks the A2A protocol. + +# Useful when you want to mount the app inside another ASGI framework +# (e.g. FastAPI), run it with a custom server, or use it in tests. + +# Args: +# compiled_graph: A compiled agentflow graph. +# agent_card: The ``AgentCard`` describing this agent. +# streaming: Whether to use ``astream`` vs ``ainvoke`` in the +# executor. +# executor_config: Optional base config forwarded to the graph +# (e.g. ``{"recursion_limit": 50}``). + +# Returns: +# A ``Starlette`` application ready to be served. +# """ +# a2a_apps = import_a2a_module("a2a.server.apps", "build_a2a_app") +# request_handlers = import_a2a_module("a2a.server.request_handlers", "build_a2a_app") +# tasks = import_a2a_module("a2a.server.tasks", "build_a2a_app") + +# from .executor import AgentFlowExecutor + +# executor = AgentFlowExecutor( +# compiled_graph, +# config=executor_config, +# streaming=streaming, +# ) +# handler = request_handlers.DefaultRequestHandler( +# agent_executor=executor, +# task_store=tasks.InMemoryTaskStore(), +# ) +# a2a_app = a2a_apps.A2AStarletteApplication( +# agent_card=agent_card, +# http_handler=handler, +# ) +# return a2a_app.build() + + +# # ---------------------------------------------------------------------- # +# # One-call server # +# # ---------------------------------------------------------------------- # + + +# def create_a2a_server( +# compiled_graph: CompiledGraph, +# agent_card: AgentCard, +# *, +# host: str = "127.0.0.1", +# port: int = 9999, +# streaming: bool = False, +# executor_config: dict[str, Any] | None = None, +# ) -> None: +# """Build and run an A2A server exposing the given graph. + +# This is a blocking call — it starts uvicorn and does not return until +# the server is shut down. + +# Args: +# compiled_graph: A compiled agentflow graph. +# agent_card: The ``AgentCard`` describing this agent. +# host: Bind address. +# port: Bind port. +# streaming: Whether to use ``astream`` in the executor. +# executor_config: Optional base config forwarded to the graph. +# """ +# try: +# uvicorn = __import__("uvicorn") +# except Exception as exc: +# raise RuntimeError( +# "create_a2a_server requires the optional 'uvicorn' package. " +# f"{A2A_EXTRA_INSTALL_HINT}" +# ) from exc + +# app = build_a2a_app( +# compiled_graph, +# agent_card, +# streaming=streaming, +# executor_config=executor_config, +# ) +# uvicorn.run(app, host=host, port=port) diff --git a/eval_reports/s-file_20260610_231346.html b/eval_reports/s-file_20260610_231346.html new file mode 100644 index 0000000..eb1816c --- /dev/null +++ b/eval_reports/s-file_20260610_231346.html @@ -0,0 +1,954 @@ + + + + + + s-file + + + +
+ +
+
+
+ +
+ AgentFlow + Evaluation Report +
+ s-file +
+
+ Generated: 2026-06-10 23:13:46 + + 📖 Docs + + + 💾 GitHub + + +
+
+
+ +
+
+
📋
+
1
+
Total Cases
+
+
+
✅
+
1
+
Passed
+
+
+
❌
+
0
+
Failed
+
+
+
⚠️
+
0
+
Errors
+
+
+
📈
+
100%
+
Pass Rate
+
+
+
+
+
+
⏱️
+
0.00s
+
Duration
+
+ +
+ +
+
+

📊 Criterion Breakdown

+
+
+
+

🎯 Score by Case

+
+
+
+ +
+
+ + + + + +
+
+
+
+ ► +
✓
+ c1 + Score: 0.00 + 0.00s +
+
+
+
+
+
+ + + +
+ + + \ No newline at end of file diff --git a/eval_reports/s-file_20260610_231346.json b/eval_reports/s-file_20260610_231346.json new file mode 100644 index 0000000..bd48aaa --- /dev/null +++ b/eval_reports/s-file_20260610_231346.json @@ -0,0 +1,123 @@ +{ + "eval_set_id": "s-file", + "eval_set_name": "", + "results": [ + { + "eval_id": "c1", + "name": "", + "passed": true, + "criterion_results": [], + "actual_trajectory": [], + "actual_tool_calls": [], + "actual_response": "", + "messages": [], + "node_responses": [], + "node_visits": [], + "duration_seconds": 0.0, + "error": null, + "metadata": {}, + "turn_results": [], + "token_usage": { + "input_tokens": 0, + "output_tokens": 0, + "cache_read_tokens": 0, + "cache_creation_tokens": 0, + "total_tokens": 0 + }, + "agent_token_usage": { + "input_tokens": 0, + "output_tokens": 0, + "cache_read_tokens": 0, + "cache_creation_tokens": 0, + "total_tokens": 0 + }, + "node_details": [] + } + ], + "summary": { + "total_cases": 1, + "passed_cases": 1, + "failed_cases": 0, + "error_cases": 0, + "pass_rate": 1.0, + "avg_duration_seconds": 0.0, + "total_duration_seconds": 0.0, + "criterion_stats": {}, + "total_token_usage": { + "input_tokens": 0, + "output_tokens": 0, + "cache_read_tokens": 0, + "cache_creation_tokens": 0, + "total_tokens": 0 + }, + "per_case_token_usage": { + "c1": { + "input_tokens": 0, + "output_tokens": 0, + "cache_read_tokens": 0, + "cache_creation_tokens": 0, + "total_tokens": 0 + } + }, + "avg_tokens_per_case": 0.0 + }, + "config_used": { + "criteria": { + "tool_name_match": null, + "trajectory": { + "threshold": 1.0, + "match_type": "EXACT", + "judge_model": "gemini-2.5-flash", + "num_samples": 3, + "rubrics": [], + "keywords": [], + "check_args": false, + "enabled": true, + "api_style": "responses" + }, + "node_order": null, + "response_match": { + "threshold": 0.8, + "match_type": "EXACT", + "judge_model": "gemini-2.5-flash", + "num_samples": 3, + "rubrics": [], + "keywords": [], + "check_args": false, + "enabled": true, + "api_style": "responses" + }, + "rouge_match": null, + "contains_keywords": null, + "llm_judge": null, + "rubric_based": null, + "factual_accuracy": null, + "hallucination": null, + "safety": null, + "simulation_goals": null + }, + "user_simulator_config": null, + "parallel": false, + "max_concurrency": 4, + "timeout": 300.0, + "verbose": false, + "mock_mode": false, + "reporter": { + "enabled": true, + "output_dir": "eval_reports", + "console": true, + "json_report": true, + "html": true, + "junit_xml": false, + "verbose": true, + "include_details": true, + "include_trajectory": true, + "include_node_responses": true, + "include_actual_response": true, + "include_tool_call_details": true, + "timestamp_files": true + } + }, + "timestamp": 1781111626.7762134, + "metadata": {} +} \ No newline at end of file diff --git a/eval_reports/s-file_20260610_232729.html b/eval_reports/s-file_20260610_232729.html new file mode 100644 index 0000000..f8d0069 --- /dev/null +++ b/eval_reports/s-file_20260610_232729.html @@ -0,0 +1,954 @@ + + + + + + s-file + + + +
+ +
+
+
+ +
+ AgentFlow + Evaluation Report +
+ s-file +
+
+ Generated: 2026-06-10 23:27:29 + + 📖 Docs + + + 💾 GitHub + + +
+
+
+ +
+
+
📋
+
1
+
Total Cases
+
+
+
✅
+
1
+
Passed
+
+
+
❌
+
0
+
Failed
+
+
+
⚠️
+
0
+
Errors
+
+
+
📈
+
100%
+
Pass Rate
+
+
+
+
+
+
⏱️
+
0.00s
+
Duration
+
+ +
+ +
+
+

📊 Criterion Breakdown

+
+
+
+

🎯 Score by Case

+
+
+
+ +
+
+ + + + + +
+
+
+
+ ► +
✓
+ c1 + Score: 0.00 + 0.00s +
+
+
+
+
+
+ + + +
+ + + \ No newline at end of file diff --git a/eval_reports/s-file_20260610_232729.json b/eval_reports/s-file_20260610_232729.json new file mode 100644 index 0000000..e4756f6 --- /dev/null +++ b/eval_reports/s-file_20260610_232729.json @@ -0,0 +1,123 @@ +{ + "eval_set_id": "s-file", + "eval_set_name": "", + "results": [ + { + "eval_id": "c1", + "name": "", + "passed": true, + "criterion_results": [], + "actual_trajectory": [], + "actual_tool_calls": [], + "actual_response": "", + "messages": [], + "node_responses": [], + "node_visits": [], + "duration_seconds": 0.0, + "error": null, + "metadata": {}, + "turn_results": [], + "token_usage": { + "input_tokens": 0, + "output_tokens": 0, + "cache_read_tokens": 0, + "cache_creation_tokens": 0, + "total_tokens": 0 + }, + "agent_token_usage": { + "input_tokens": 0, + "output_tokens": 0, + "cache_read_tokens": 0, + "cache_creation_tokens": 0, + "total_tokens": 0 + }, + "node_details": [] + } + ], + "summary": { + "total_cases": 1, + "passed_cases": 1, + "failed_cases": 0, + "error_cases": 0, + "pass_rate": 1.0, + "avg_duration_seconds": 0.0, + "total_duration_seconds": 0.0, + "criterion_stats": {}, + "total_token_usage": { + "input_tokens": 0, + "output_tokens": 0, + "cache_read_tokens": 0, + "cache_creation_tokens": 0, + "total_tokens": 0 + }, + "per_case_token_usage": { + "c1": { + "input_tokens": 0, + "output_tokens": 0, + "cache_read_tokens": 0, + "cache_creation_tokens": 0, + "total_tokens": 0 + } + }, + "avg_tokens_per_case": 0.0 + }, + "config_used": { + "criteria": { + "tool_name_match": null, + "trajectory": { + "threshold": 1.0, + "match_type": "EXACT", + "judge_model": "gemini-2.5-flash", + "num_samples": 3, + "rubrics": [], + "keywords": [], + "check_args": false, + "enabled": true, + "api_style": "responses" + }, + "node_order": null, + "response_match": { + "threshold": 0.8, + "match_type": "EXACT", + "judge_model": "gemini-2.5-flash", + "num_samples": 3, + "rubrics": [], + "keywords": [], + "check_args": false, + "enabled": true, + "api_style": "responses" + }, + "rouge_match": null, + "contains_keywords": null, + "llm_judge": null, + "rubric_based": null, + "factual_accuracy": null, + "hallucination": null, + "safety": null, + "simulation_goals": null + }, + "user_simulator_config": null, + "parallel": false, + "max_concurrency": 4, + "timeout": 300.0, + "verbose": false, + "mock_mode": false, + "reporter": { + "enabled": true, + "output_dir": "eval_reports", + "console": true, + "json_report": true, + "html": true, + "junit_xml": false, + "verbose": true, + "include_details": true, + "include_trajectory": true, + "include_node_responses": true, + "include_actual_response": true, + "include_tool_call_details": true, + "timestamp_files": true + } + }, + "timestamp": 1781112449.644461, + "metadata": {} +} \ No newline at end of file diff --git a/eval_reports/s-file_20260610_233004.html b/eval_reports/s-file_20260610_233004.html new file mode 100644 index 0000000..f099f4a --- /dev/null +++ b/eval_reports/s-file_20260610_233004.html @@ -0,0 +1,954 @@ + + + + + + s-file + + + +
+ +
+
+
+ +
+ AgentFlow + Evaluation Report +
+ s-file +
+
+ Generated: 2026-06-10 23:30:04 + + 📖 Docs + + + 💾 GitHub + + +
+
+
+ +
+
+
📋
+
1
+
Total Cases
+
+
+
✅
+
1
+
Passed
+
+
+
❌
+
0
+
Failed
+
+
+
⚠️
+
0
+
Errors
+
+
+
📈
+
100%
+
Pass Rate
+
+
+
+
+
+
⏱️
+
0.00s
+
Duration
+
+ +
+ +
+
+

📊 Criterion Breakdown

+
+
+
+

🎯 Score by Case

+
+
+
+ +
+
+ + + + + +
+
+
+
+ ► +
✓
+ c1 + Score: 0.00 + 0.00s +
+
+
+
+
+
+ + + +
+ + + \ No newline at end of file diff --git a/eval_reports/s-file_20260610_233004.json b/eval_reports/s-file_20260610_233004.json new file mode 100644 index 0000000..33b19b9 --- /dev/null +++ b/eval_reports/s-file_20260610_233004.json @@ -0,0 +1,123 @@ +{ + "eval_set_id": "s-file", + "eval_set_name": "", + "results": [ + { + "eval_id": "c1", + "name": "", + "passed": true, + "criterion_results": [], + "actual_trajectory": [], + "actual_tool_calls": [], + "actual_response": "", + "messages": [], + "node_responses": [], + "node_visits": [], + "duration_seconds": 0.0, + "error": null, + "metadata": {}, + "turn_results": [], + "token_usage": { + "input_tokens": 0, + "output_tokens": 0, + "cache_read_tokens": 0, + "cache_creation_tokens": 0, + "total_tokens": 0 + }, + "agent_token_usage": { + "input_tokens": 0, + "output_tokens": 0, + "cache_read_tokens": 0, + "cache_creation_tokens": 0, + "total_tokens": 0 + }, + "node_details": [] + } + ], + "summary": { + "total_cases": 1, + "passed_cases": 1, + "failed_cases": 0, + "error_cases": 0, + "pass_rate": 1.0, + "avg_duration_seconds": 0.0, + "total_duration_seconds": 0.0, + "criterion_stats": {}, + "total_token_usage": { + "input_tokens": 0, + "output_tokens": 0, + "cache_read_tokens": 0, + "cache_creation_tokens": 0, + "total_tokens": 0 + }, + "per_case_token_usage": { + "c1": { + "input_tokens": 0, + "output_tokens": 0, + "cache_read_tokens": 0, + "cache_creation_tokens": 0, + "total_tokens": 0 + } + }, + "avg_tokens_per_case": 0.0 + }, + "config_used": { + "criteria": { + "tool_name_match": null, + "trajectory": { + "threshold": 1.0, + "match_type": "EXACT", + "judge_model": "gemini-2.5-flash", + "num_samples": 3, + "rubrics": [], + "keywords": [], + "check_args": false, + "enabled": true, + "api_style": "responses" + }, + "node_order": null, + "response_match": { + "threshold": 0.8, + "match_type": "EXACT", + "judge_model": "gemini-2.5-flash", + "num_samples": 3, + "rubrics": [], + "keywords": [], + "check_args": false, + "enabled": true, + "api_style": "responses" + }, + "rouge_match": null, + "contains_keywords": null, + "llm_judge": null, + "rubric_based": null, + "factual_accuracy": null, + "hallucination": null, + "safety": null, + "simulation_goals": null + }, + "user_simulator_config": null, + "parallel": false, + "max_concurrency": 4, + "timeout": 300.0, + "verbose": false, + "mock_mode": false, + "reporter": { + "enabled": true, + "output_dir": "eval_reports", + "console": true, + "json_report": true, + "html": true, + "junit_xml": false, + "verbose": true, + "include_details": true, + "include_trajectory": true, + "include_node_responses": true, + "include_actual_response": true, + "include_tool_call_details": true, + "timestamp_files": true + } + }, + "timestamp": 1781112604.7925491, + "metadata": {} +} \ No newline at end of file diff --git a/eval_reports/s-file_20260610_234128.html b/eval_reports/s-file_20260610_234128.html new file mode 100644 index 0000000..e82b5cb --- /dev/null +++ b/eval_reports/s-file_20260610_234128.html @@ -0,0 +1,954 @@ + + + + + + s-file + + + +
+ +
+
+
+ +
+ AgentFlow + Evaluation Report +
+ s-file +
+
+ Generated: 2026-06-10 23:41:28 + + 📖 Docs + + + 💾 GitHub + + +
+
+
+ +
+
+
📋
+
1
+
Total Cases
+
+
+
✅
+
1
+
Passed
+
+
+
❌
+
0
+
Failed
+
+
+
⚠️
+
0
+
Errors
+
+
+
📈
+
100%
+
Pass Rate
+
+
+
+
+
+
⏱️
+
0.00s
+
Duration
+
+ +
+ +
+
+

📊 Criterion Breakdown

+
+
+
+

🎯 Score by Case

+
+
+
+ +
+
+ + + + + +
+
+
+
+ ► +
✓
+ c1 + Score: 0.00 + 0.00s +
+
+
+
+
+
+ + + +
+ + + \ No newline at end of file diff --git a/eval_reports/s-file_20260610_234128.json b/eval_reports/s-file_20260610_234128.json new file mode 100644 index 0000000..dd3a2ad --- /dev/null +++ b/eval_reports/s-file_20260610_234128.json @@ -0,0 +1,123 @@ +{ + "eval_set_id": "s-file", + "eval_set_name": "", + "results": [ + { + "eval_id": "c1", + "name": "", + "passed": true, + "criterion_results": [], + "actual_trajectory": [], + "actual_tool_calls": [], + "actual_response": "", + "messages": [], + "node_responses": [], + "node_visits": [], + "duration_seconds": 0.0, + "error": null, + "metadata": {}, + "turn_results": [], + "token_usage": { + "input_tokens": 0, + "output_tokens": 0, + "cache_read_tokens": 0, + "cache_creation_tokens": 0, + "total_tokens": 0 + }, + "agent_token_usage": { + "input_tokens": 0, + "output_tokens": 0, + "cache_read_tokens": 0, + "cache_creation_tokens": 0, + "total_tokens": 0 + }, + "node_details": [] + } + ], + "summary": { + "total_cases": 1, + "passed_cases": 1, + "failed_cases": 0, + "error_cases": 0, + "pass_rate": 1.0, + "avg_duration_seconds": 0.0, + "total_duration_seconds": 0.0, + "criterion_stats": {}, + "total_token_usage": { + "input_tokens": 0, + "output_tokens": 0, + "cache_read_tokens": 0, + "cache_creation_tokens": 0, + "total_tokens": 0 + }, + "per_case_token_usage": { + "c1": { + "input_tokens": 0, + "output_tokens": 0, + "cache_read_tokens": 0, + "cache_creation_tokens": 0, + "total_tokens": 0 + } + }, + "avg_tokens_per_case": 0.0 + }, + "config_used": { + "criteria": { + "tool_name_match": null, + "trajectory": { + "threshold": 1.0, + "match_type": "EXACT", + "judge_model": "gemini-2.5-flash", + "num_samples": 3, + "rubrics": [], + "keywords": [], + "check_args": false, + "enabled": true, + "api_style": "responses" + }, + "node_order": null, + "response_match": { + "threshold": 0.8, + "match_type": "EXACT", + "judge_model": "gemini-2.5-flash", + "num_samples": 3, + "rubrics": [], + "keywords": [], + "check_args": false, + "enabled": true, + "api_style": "responses" + }, + "rouge_match": null, + "contains_keywords": null, + "llm_judge": null, + "rubric_based": null, + "factual_accuracy": null, + "hallucination": null, + "safety": null, + "simulation_goals": null + }, + "user_simulator_config": null, + "parallel": false, + "max_concurrency": 4, + "timeout": 300.0, + "verbose": false, + "mock_mode": false, + "reporter": { + "enabled": true, + "output_dir": "eval_reports", + "console": true, + "json_report": true, + "html": true, + "junit_xml": false, + "verbose": true, + "include_details": true, + "include_trajectory": true, + "include_node_responses": true, + "include_actual_response": true, + "include_tool_call_details": true, + "timestamp_files": true + } + }, + "timestamp": 1781113288.1752944, + "metadata": {} +} \ No newline at end of file diff --git a/eval_reports/s-file_20260611_000420.html b/eval_reports/s-file_20260611_000420.html new file mode 100644 index 0000000..9c2aa21 --- /dev/null +++ b/eval_reports/s-file_20260611_000420.html @@ -0,0 +1,954 @@ + + + + + + s-file + + + +
+ +
+
+
+ +
+ AgentFlow + Evaluation Report +
+ s-file +
+
+ Generated: 2026-06-11 00:04:20 + + 📖 Docs + + + 💾 GitHub + + +
+
+
+ +
+
+
📋
+
1
+
Total Cases
+
+
+
✅
+
1
+
Passed
+
+
+
❌
+
0
+
Failed
+
+
+
⚠️
+
0
+
Errors
+
+
+
📈
+
100%
+
Pass Rate
+
+
+
+
+
+
⏱️
+
0.00s
+
Duration
+
+ +
+ +
+
+

📊 Criterion Breakdown

+
+
+
+

🎯 Score by Case

+
+
+
+ +
+
+ + + + + +
+
+
+
+ ► +
✓
+ c1 + Score: 0.00 + 0.00s +
+
+
+
+
+
+ + + +
+ + + \ No newline at end of file diff --git a/eval_reports/s-file_20260611_000420.json b/eval_reports/s-file_20260611_000420.json new file mode 100644 index 0000000..1310bb8 --- /dev/null +++ b/eval_reports/s-file_20260611_000420.json @@ -0,0 +1,123 @@ +{ + "eval_set_id": "s-file", + "eval_set_name": "", + "results": [ + { + "eval_id": "c1", + "name": "", + "passed": true, + "criterion_results": [], + "actual_trajectory": [], + "actual_tool_calls": [], + "actual_response": "", + "messages": [], + "node_responses": [], + "node_visits": [], + "duration_seconds": 0.0, + "error": null, + "metadata": {}, + "turn_results": [], + "token_usage": { + "input_tokens": 0, + "output_tokens": 0, + "cache_read_tokens": 0, + "cache_creation_tokens": 0, + "total_tokens": 0 + }, + "agent_token_usage": { + "input_tokens": 0, + "output_tokens": 0, + "cache_read_tokens": 0, + "cache_creation_tokens": 0, + "total_tokens": 0 + }, + "node_details": [] + } + ], + "summary": { + "total_cases": 1, + "passed_cases": 1, + "failed_cases": 0, + "error_cases": 0, + "pass_rate": 1.0, + "avg_duration_seconds": 0.0, + "total_duration_seconds": 0.0, + "criterion_stats": {}, + "total_token_usage": { + "input_tokens": 0, + "output_tokens": 0, + "cache_read_tokens": 0, + "cache_creation_tokens": 0, + "total_tokens": 0 + }, + "per_case_token_usage": { + "c1": { + "input_tokens": 0, + "output_tokens": 0, + "cache_read_tokens": 0, + "cache_creation_tokens": 0, + "total_tokens": 0 + } + }, + "avg_tokens_per_case": 0.0 + }, + "config_used": { + "criteria": { + "tool_name_match": null, + "trajectory": { + "threshold": 1.0, + "match_type": "EXACT", + "judge_model": "gemini-2.5-flash", + "num_samples": 3, + "rubrics": [], + "keywords": [], + "check_args": false, + "enabled": true, + "api_style": "responses" + }, + "node_order": null, + "response_match": { + "threshold": 0.8, + "match_type": "EXACT", + "judge_model": "gemini-2.5-flash", + "num_samples": 3, + "rubrics": [], + "keywords": [], + "check_args": false, + "enabled": true, + "api_style": "responses" + }, + "rouge_match": null, + "contains_keywords": null, + "llm_judge": null, + "rubric_based": null, + "factual_accuracy": null, + "hallucination": null, + "safety": null, + "simulation_goals": null + }, + "user_simulator_config": null, + "parallel": false, + "max_concurrency": 4, + "timeout": 300.0, + "verbose": false, + "mock_mode": false, + "reporter": { + "enabled": true, + "output_dir": "eval_reports", + "console": true, + "json_report": true, + "html": true, + "junit_xml": false, + "verbose": true, + "include_details": true, + "include_trajectory": true, + "include_node_responses": true, + "include_actual_response": true, + "include_tool_call_details": true, + "timestamp_files": true + } + }, + "timestamp": 1781114660.4083369, + "metadata": {} +} \ No newline at end of file diff --git a/eval_reports/s-file_20260611_000529.html b/eval_reports/s-file_20260611_000529.html new file mode 100644 index 0000000..bbce957 --- /dev/null +++ b/eval_reports/s-file_20260611_000529.html @@ -0,0 +1,954 @@ + + + + + + s-file + + + +
+ +
+
+
+ +
+ AgentFlow + Evaluation Report +
+ s-file +
+
+ Generated: 2026-06-11 00:05:29 + + 📖 Docs + + + 💾 GitHub + + +
+
+
+ +
+
+
📋
+
1
+
Total Cases
+
+
+
✅
+
1
+
Passed
+
+
+
❌
+
0
+
Failed
+
+
+
⚠️
+
0
+
Errors
+
+
+
📈
+
100%
+
Pass Rate
+
+
+
+
+
+
⏱️
+
0.00s
+
Duration
+
+ +
+ +
+
+

📊 Criterion Breakdown

+
+
+
+

🎯 Score by Case

+
+
+
+ +
+
+ + + + + +
+
+
+
+ ► +
✓
+ c1 + Score: 0.00 + 0.00s +
+
+
+
+
+
+ + + +
+ + + \ No newline at end of file diff --git a/eval_reports/s-file_20260611_000529.json b/eval_reports/s-file_20260611_000529.json new file mode 100644 index 0000000..694b176 --- /dev/null +++ b/eval_reports/s-file_20260611_000529.json @@ -0,0 +1,123 @@ +{ + "eval_set_id": "s-file", + "eval_set_name": "", + "results": [ + { + "eval_id": "c1", + "name": "", + "passed": true, + "criterion_results": [], + "actual_trajectory": [], + "actual_tool_calls": [], + "actual_response": "", + "messages": [], + "node_responses": [], + "node_visits": [], + "duration_seconds": 0.0, + "error": null, + "metadata": {}, + "turn_results": [], + "token_usage": { + "input_tokens": 0, + "output_tokens": 0, + "cache_read_tokens": 0, + "cache_creation_tokens": 0, + "total_tokens": 0 + }, + "agent_token_usage": { + "input_tokens": 0, + "output_tokens": 0, + "cache_read_tokens": 0, + "cache_creation_tokens": 0, + "total_tokens": 0 + }, + "node_details": [] + } + ], + "summary": { + "total_cases": 1, + "passed_cases": 1, + "failed_cases": 0, + "error_cases": 0, + "pass_rate": 1.0, + "avg_duration_seconds": 0.0, + "total_duration_seconds": 0.0, + "criterion_stats": {}, + "total_token_usage": { + "input_tokens": 0, + "output_tokens": 0, + "cache_read_tokens": 0, + "cache_creation_tokens": 0, + "total_tokens": 0 + }, + "per_case_token_usage": { + "c1": { + "input_tokens": 0, + "output_tokens": 0, + "cache_read_tokens": 0, + "cache_creation_tokens": 0, + "total_tokens": 0 + } + }, + "avg_tokens_per_case": 0.0 + }, + "config_used": { + "criteria": { + "tool_name_match": null, + "trajectory": { + "threshold": 1.0, + "match_type": "EXACT", + "judge_model": "gemini-2.5-flash", + "num_samples": 3, + "rubrics": [], + "keywords": [], + "check_args": false, + "enabled": true, + "api_style": "responses" + }, + "node_order": null, + "response_match": { + "threshold": 0.8, + "match_type": "EXACT", + "judge_model": "gemini-2.5-flash", + "num_samples": 3, + "rubrics": [], + "keywords": [], + "check_args": false, + "enabled": true, + "api_style": "responses" + }, + "rouge_match": null, + "contains_keywords": null, + "llm_judge": null, + "rubric_based": null, + "factual_accuracy": null, + "hallucination": null, + "safety": null, + "simulation_goals": null + }, + "user_simulator_config": null, + "parallel": false, + "max_concurrency": 4, + "timeout": 300.0, + "verbose": false, + "mock_mode": false, + "reporter": { + "enabled": true, + "output_dir": "eval_reports", + "console": true, + "json_report": true, + "html": true, + "junit_xml": false, + "verbose": true, + "include_details": true, + "include_trajectory": true, + "include_node_responses": true, + "include_actual_response": true, + "include_tool_call_details": true, + "timestamp_files": true + } + }, + "timestamp": 1781114729.9052248, + "metadata": {} +} \ No newline at end of file diff --git a/eval_reports/s-file_20260611_001112.html b/eval_reports/s-file_20260611_001112.html new file mode 100644 index 0000000..642c8f8 --- /dev/null +++ b/eval_reports/s-file_20260611_001112.html @@ -0,0 +1,954 @@ + + + + + + s-file + + + +
+ +
+
+
+ +
+ AgentFlow + Evaluation Report +
+ s-file +
+
+ Generated: 2026-06-11 00:11:12 + + 📖 Docs + + + 💾 GitHub + + +
+
+
+ +
+
+
📋
+
1
+
Total Cases
+
+
+
✅
+
1
+
Passed
+
+
+
❌
+
0
+
Failed
+
+
+
⚠️
+
0
+
Errors
+
+
+
📈
+
100%
+
Pass Rate
+
+
+
+
+
+
⏱️
+
0.00s
+
Duration
+
+ +
+ +
+
+

📊 Criterion Breakdown

+
+
+
+

🎯 Score by Case

+
+
+
+ +
+
+ + + + + +
+
+
+
+ ► +
✓
+ c1 + Score: 0.00 + 0.00s +
+
+
+
+
+
+ + + +
+ + + \ No newline at end of file diff --git a/eval_reports/s-file_20260611_001112.json b/eval_reports/s-file_20260611_001112.json new file mode 100644 index 0000000..c7174a5 --- /dev/null +++ b/eval_reports/s-file_20260611_001112.json @@ -0,0 +1,123 @@ +{ + "eval_set_id": "s-file", + "eval_set_name": "", + "results": [ + { + "eval_id": "c1", + "name": "", + "passed": true, + "criterion_results": [], + "actual_trajectory": [], + "actual_tool_calls": [], + "actual_response": "", + "messages": [], + "node_responses": [], + "node_visits": [], + "duration_seconds": 0.0, + "error": null, + "metadata": {}, + "turn_results": [], + "token_usage": { + "input_tokens": 0, + "output_tokens": 0, + "cache_read_tokens": 0, + "cache_creation_tokens": 0, + "total_tokens": 0 + }, + "agent_token_usage": { + "input_tokens": 0, + "output_tokens": 0, + "cache_read_tokens": 0, + "cache_creation_tokens": 0, + "total_tokens": 0 + }, + "node_details": [] + } + ], + "summary": { + "total_cases": 1, + "passed_cases": 1, + "failed_cases": 0, + "error_cases": 0, + "pass_rate": 1.0, + "avg_duration_seconds": 0.0, + "total_duration_seconds": 0.0, + "criterion_stats": {}, + "total_token_usage": { + "input_tokens": 0, + "output_tokens": 0, + "cache_read_tokens": 0, + "cache_creation_tokens": 0, + "total_tokens": 0 + }, + "per_case_token_usage": { + "c1": { + "input_tokens": 0, + "output_tokens": 0, + "cache_read_tokens": 0, + "cache_creation_tokens": 0, + "total_tokens": 0 + } + }, + "avg_tokens_per_case": 0.0 + }, + "config_used": { + "criteria": { + "tool_name_match": null, + "trajectory": { + "threshold": 1.0, + "match_type": "EXACT", + "judge_model": "gemini-2.5-flash", + "num_samples": 3, + "rubrics": [], + "keywords": [], + "check_args": false, + "enabled": true, + "api_style": "responses" + }, + "node_order": null, + "response_match": { + "threshold": 0.8, + "match_type": "EXACT", + "judge_model": "gemini-2.5-flash", + "num_samples": 3, + "rubrics": [], + "keywords": [], + "check_args": false, + "enabled": true, + "api_style": "responses" + }, + "rouge_match": null, + "contains_keywords": null, + "llm_judge": null, + "rubric_based": null, + "factual_accuracy": null, + "hallucination": null, + "safety": null, + "simulation_goals": null + }, + "user_simulator_config": null, + "parallel": false, + "max_concurrency": 4, + "timeout": 300.0, + "verbose": false, + "mock_mode": false, + "reporter": { + "enabled": true, + "output_dir": "eval_reports", + "console": true, + "json_report": true, + "html": true, + "junit_xml": false, + "verbose": true, + "include_details": true, + "include_trajectory": true, + "include_node_responses": true, + "include_actual_response": true, + "include_tool_call_details": true, + "timestamp_files": true + } + }, + "timestamp": 1781115072.5616362, + "metadata": {} +} \ No newline at end of file diff --git a/eval_reports/s-file_20260611_001220.html b/eval_reports/s-file_20260611_001220.html new file mode 100644 index 0000000..a494286 --- /dev/null +++ b/eval_reports/s-file_20260611_001220.html @@ -0,0 +1,954 @@ + + + + + + s-file + + + +
+ +
+
+
+ +
+ AgentFlow + Evaluation Report +
+ s-file +
+
+ Generated: 2026-06-11 00:12:20 + + 📖 Docs + + + 💾 GitHub + + +
+
+
+ +
+
+
📋
+
1
+
Total Cases
+
+
+
✅
+
1
+
Passed
+
+
+
❌
+
0
+
Failed
+
+
+
⚠️
+
0
+
Errors
+
+
+
📈
+
100%
+
Pass Rate
+
+
+
+
+
+
⏱️
+
0.00s
+
Duration
+
+ +
+ +
+
+

📊 Criterion Breakdown

+
+
+
+

🎯 Score by Case

+
+
+
+ +
+
+ + + + + +
+
+
+
+ ► +
✓
+ c1 + Score: 0.00 + 0.00s +
+
+
+
+
+
+ + + +
+ + + \ No newline at end of file diff --git a/eval_reports/s-file_20260611_001220.json b/eval_reports/s-file_20260611_001220.json new file mode 100644 index 0000000..74a23de --- /dev/null +++ b/eval_reports/s-file_20260611_001220.json @@ -0,0 +1,123 @@ +{ + "eval_set_id": "s-file", + "eval_set_name": "", + "results": [ + { + "eval_id": "c1", + "name": "", + "passed": true, + "criterion_results": [], + "actual_trajectory": [], + "actual_tool_calls": [], + "actual_response": "", + "messages": [], + "node_responses": [], + "node_visits": [], + "duration_seconds": 0.0, + "error": null, + "metadata": {}, + "turn_results": [], + "token_usage": { + "input_tokens": 0, + "output_tokens": 0, + "cache_read_tokens": 0, + "cache_creation_tokens": 0, + "total_tokens": 0 + }, + "agent_token_usage": { + "input_tokens": 0, + "output_tokens": 0, + "cache_read_tokens": 0, + "cache_creation_tokens": 0, + "total_tokens": 0 + }, + "node_details": [] + } + ], + "summary": { + "total_cases": 1, + "passed_cases": 1, + "failed_cases": 0, + "error_cases": 0, + "pass_rate": 1.0, + "avg_duration_seconds": 0.0, + "total_duration_seconds": 0.0, + "criterion_stats": {}, + "total_token_usage": { + "input_tokens": 0, + "output_tokens": 0, + "cache_read_tokens": 0, + "cache_creation_tokens": 0, + "total_tokens": 0 + }, + "per_case_token_usage": { + "c1": { + "input_tokens": 0, + "output_tokens": 0, + "cache_read_tokens": 0, + "cache_creation_tokens": 0, + "total_tokens": 0 + } + }, + "avg_tokens_per_case": 0.0 + }, + "config_used": { + "criteria": { + "tool_name_match": null, + "trajectory": { + "threshold": 1.0, + "match_type": "EXACT", + "judge_model": "gemini-2.5-flash", + "num_samples": 3, + "rubrics": [], + "keywords": [], + "check_args": false, + "enabled": true, + "api_style": "responses" + }, + "node_order": null, + "response_match": { + "threshold": 0.8, + "match_type": "EXACT", + "judge_model": "gemini-2.5-flash", + "num_samples": 3, + "rubrics": [], + "keywords": [], + "check_args": false, + "enabled": true, + "api_style": "responses" + }, + "rouge_match": null, + "contains_keywords": null, + "llm_judge": null, + "rubric_based": null, + "factual_accuracy": null, + "hallucination": null, + "safety": null, + "simulation_goals": null + }, + "user_simulator_config": null, + "parallel": false, + "max_concurrency": 4, + "timeout": 300.0, + "verbose": false, + "mock_mode": false, + "reporter": { + "enabled": true, + "output_dir": "eval_reports", + "console": true, + "json_report": true, + "html": true, + "junit_xml": false, + "verbose": true, + "include_details": true, + "include_trajectory": true, + "include_node_responses": true, + "include_actual_response": true, + "include_tool_call_details": true, + "timestamp_files": true + } + }, + "timestamp": 1781115140.0770233, + "metadata": {} +} \ No newline at end of file diff --git a/pyproject.toml b/pyproject.toml index 083dd63..76a4b0b 100644 --- a/pyproject.toml +++ b/pyproject.toml @@ -1,6 +1,6 @@ [project] name = "10xscale-agentflow" -version = "0.7.5.0" +version = "0.7.5.1" description = "Production-grade Python framework for building, orchestrating, and deploying multi-agent LLM systems. A simpler, batteries-included alternative to LangGraph, CrewAI, and AutoGen with graph-based workflows, durable state, native MCP support, and provider-agnostic LLM integration (OpenAI, Google GenAI, Anthropic)." readme = "README.md" keywords = [ diff --git a/tests/adapters/test_openai_converter.py b/tests/adapters/test_openai_converter.py index eecdd8a..a5d7ac4 100644 --- a/tests/adapters/test_openai_converter.py +++ b/tests/adapters/test_openai_converter.py @@ -438,3 +438,175 @@ async def test_no_token_details_defaults_to_zero(self): assert msg.usages.reasoning_tokens == 0 assert msg.usages.cache_read_input_tokens == 0 assert msg.usages.cache_creation_input_tokens == 0 + + +@pytest.fixture +def converter(): + return OpenAIConverter() + + +@pytest.mark.asyncio +async def test_openai_converter_import_errors(monkeypatch): + monkeypatch.setattr("agentflow.runtime.adapters.llm.openai_converter.HAS_OPENAI", False) + converter = OpenAIConverter() + + with pytest.raises(ImportError, match="openai is not installed"): + await converter.convert_response(Mock()) + + with pytest.raises(ImportError, match="openai is not installed"): + async for _ in converter.convert_streaming_response({}, "node", Mock()): + pass + + +def test_extract_audio_block_exceptions(converter): + assert converter._extract_audio_block({"transcript": "hi"}) is None + assert converter._extract_audio_block(object()) is None + + +def test_extract_image_blocks_various_types(converter): + img = "https://example.com/single.png" + blocks = converter._extract_image_blocks(img) + assert len(blocks) == 1 + assert blocks[0].media.url == img + + with patch("agentflow.runtime.adapters.llm.openai_converter.isinstance", side_effect=TypeError("mock error")): + blocks2 = converter._extract_image_blocks("url") + assert blocks2 == [] + + +@pytest.mark.asyncio +async def test_streaming_conversion(converter): + class MockChunk: + def __init__(self, id, content=None, reasoning=None, tool_calls=None): + self.id = id + self.model = "gpt-4o" + + delta_obj = type("Delta", (), { + "content": content, + "reasoning_content": reasoning, + "tool_calls": tool_calls, + "audio": None, + "images": None + }) + choice = type("Choice", (), { + "delta": delta_obj + }) + self.choices = [choice] + + chunk1 = MockChunk("chat-1", reasoning="Thinking...") + chunk2 = MockChunk("chat-1", content="Hello ") + chunk3 = MockChunk("chat-1", content="world!") + + tool_call_mock = type("ToolCall", (), { + "id": "tc-123", + "type": "function", + "function": type("Func", (), { + "name": "calc", + "arguments": '{"x": 1}' + }) + }) + chunk4 = MockChunk("chat-1", tool_calls=[tool_call_mock]) + + async def mock_async_stream(): + yield chunk1 + yield chunk2 + yield chunk3 + yield chunk4 + + messages = [] + async for msg in converter.convert_streaming_response({}, "my_node", mock_async_stream()): + messages.append(msg) + + assert len(messages) == 5 + assert messages[0].reasoning == "Thinking..." + assert messages[1].content[0].text == "Hello " + assert messages[2].content[0].text == "world!" + assert messages[3].tools_calls[0]["id"] == "tc-123" + + final_msg = messages[-1] + assert final_msg.delta is False + assert final_msg.reasoning == "Thinking..." + assert final_msg.content[0].text == "Hello world!" + assert final_msg.content[1].summary == "Thinking..." + assert final_msg.content[2].name == "calc" + + +@pytest.mark.asyncio +async def test_streaming_inline_think_thought(converter): + class MockChunk: + def __init__(self, id, content): + self.id = id + delta_obj = type("Delta", (), { + "content": content, + "reasoning_content": None, + "tool_calls": None, + "audio": None, + "images": None + }) + self.choices = [type("Choice", (), {"delta": delta_obj})] + + async def mock_stream(): + yield MockChunk("chat-2", "Inline thoughtsActual content") + + messages = [] + async for msg in converter.convert_streaming_response({}, "node", mock_stream()): + messages.append(msg) + + final_msg = messages[-1] + assert final_msg.reasoning == "Inline thoughts" + assert final_msg.content[0].text == "Actual content" + + +@pytest.mark.asyncio +async def test_streaming_sync_iterator(converter): + class MockChunk: + def __init__(self, id, content): + self.id = id + delta_obj = type("Delta", (), { + "content": content, + "reasoning_content": None, + "tool_calls": None, + "audio": None, + "images": None + }) + self.choices = [type("Choice", (), {"delta": delta_obj})] + + class SyncStream: + def __init__(self, chunks): + self.chunks = chunks + def __iter__(self): + return iter(self.chunks) + + stream = SyncStream([MockChunk("chat-3", "Sync chunk")]) + messages = [] + async for msg in converter.convert_streaming_response({}, "node", stream): + messages.append(msg) + + assert len(messages) == 2 + assert messages[0].content[0].text == "Sync chunk" + + +@pytest.mark.asyncio +async def test_convert_streaming_response_chat_completion(converter): + response_data = { + "id": "chatcmpl-123", + "model": "gpt-4o", + "choices": [{"message": {"role": "assistant", "content": "Hello"}}], + "usage": {} + } + response = MockModelResponse(response_data) + messages = [] + with patch("agentflow.runtime.adapters.llm.openai_converter.ChatCompletion", MockModelResponse): + async for msg in converter.convert_streaming_response({}, "node", response): + messages.append(msg) + + assert len(messages) == 1 + assert messages[0].content[0].text == "Hello" + + +@pytest.mark.asyncio +async def test_convert_streaming_response_unsupported(converter): + with pytest.raises(Exception, match="Unsupported response type"): + async for _ in converter.convert_streaming_response({}, "node", object()): + pass + diff --git a/tests/checkpointer/test_base_checkpointer.py b/tests/checkpointer/test_base_checkpointer.py new file mode 100644 index 0000000..e4cdcfe --- /dev/null +++ b/tests/checkpointer/test_base_checkpointer.py @@ -0,0 +1,129 @@ +import pytest +from unittest.mock import AsyncMock, MagicMock +from agentflow.storage.checkpointer.base_checkpointer import BaseCheckpointer +from agentflow.core.state import AgentState, Message +from agentflow.utils.thread_info import ThreadInfo + + +class DummyCheckpointer(BaseCheckpointer): + async def asetup(self): pass + async def aput_state(self, config, state): pass + async def aget_state(self, config): pass + async def aclear_state(self, config): pass + async def aput_state_cache(self, config, state): pass + async def aget_state_cache(self, config): pass + async def aput_messages(self, config, messages, metadata=None): pass + async def aget_message(self, config, message_id): pass + async def alist_messages(self, config, search=None, offset=None, limit=None): pass + async def adelete_message(self, config, message_id): pass + async def aput_thread(self, config, thread_info): pass + async def aget_thread(self, config): pass + async def alist_threads(self, config, search=None, offset=None, limit=None): pass + async def aclean_thread(self, config): pass + async def arelease(self): pass + + +class MinimalCheckpointer(BaseCheckpointer): + async def asetup(self): pass + async def aput_state(self, config, state): pass + async def aget_state(self, config): pass + async def aclear_state(self, config): pass + async def aput_state_cache(self, config, state): pass + async def aget_state_cache(self, config): pass + async def aput_messages(self, config, messages, metadata=None): pass + async def aget_message(self, config, message_id): pass + async def alist_messages(self, config, search=None, offset=None, limit=None): pass + async def adelete_message(self, config, message_id): pass + async def aput_thread(self, config, thread_info): pass + async def aget_thread(self, config): pass + async def alist_threads(self, config, search=None, offset=None, limit=None): pass + async def aclean_thread(self, config): pass + async def arelease(self): pass + + +def test_base_checkpointer_sync_wrappers(): + cp = DummyCheckpointer() + cp.asetup = AsyncMock() + cp.aput_state = AsyncMock() + cp.aget_state = AsyncMock() + cp.aclear_state = AsyncMock() + cp.aput_state_cache = AsyncMock() + cp.aget_state_cache = AsyncMock() + cp.aput_messages = AsyncMock() + cp.aget_message = AsyncMock() + cp.alist_messages = AsyncMock() + cp.adelete_message = AsyncMock() + cp.aput_thread = AsyncMock() + cp.aget_thread = AsyncMock() + cp.alist_threads = AsyncMock() + cp.aclean_thread = AsyncMock() + cp.arelease = AsyncMock() + + config = {"thread_id": "test_thread"} + state = MagicMock(spec=AgentState) + messages = [MagicMock(spec=Message)] + thread_info = MagicMock(spec=ThreadInfo) + + # Test setup + cp.setup() + cp.asetup.assert_called_once() + + # Test state methods + cp.put_state(config, state) + cp.aput_state.assert_called_once_with(config, state) + + cp.get_state(config) + cp.aget_state.assert_called_once_with(config) + + cp.clear_state(config) + cp.aclear_state.assert_called_once_with(config) + + cp.put_state_cache(config, state) + cp.aput_state_cache.assert_called_once_with(config, state) + + cp.get_state_cache(config) + cp.aget_state_cache.assert_called_once_with(config) + + # Test cache values + cp.put_cache_value("ns", "k", "v", ttl_seconds=10) + cp.get_cache_value("ns", "k") + cp.clear_cache_value("ns", "k") + + # Test message methods + cp.put_messages(config, messages, metadata={"meta": 1}) + cp.aput_messages.assert_called_once_with(config, messages, {"meta": 1}) + + cp.get_message(config, "msg1") + cp.aget_message.assert_called_once_with(config, "msg1") + + cp.list_messages(config, search="x", offset=1, limit=5) + cp.alist_messages.assert_called_once_with(config, "x", 1, 5) + + cp.delete_message(config, "msg1") + cp.adelete_message.assert_called_once_with(config, "msg1") + + # Test thread methods + cp.put_thread(config, thread_info) + cp.aput_thread.assert_called_once_with(config, thread_info) + + cp.get_thread(config) + cp.aget_thread.assert_called_once_with(config) + + cp.list_threads(config, search="x", offset=1, limit=5) + cp.alist_threads.assert_called_once_with(config, "x", 1, 5) + + cp.clean_thread(config) + cp.aclean_thread.assert_called_once_with(config) + + # Test release + cp.release() + cp.arelease.assert_called_once() + + +@pytest.mark.asyncio +async def test_base_checkpointer_default_cache_methods(): + cp = MinimalCheckpointer() + assert await cp.aput_cache_value("ns", "k", "v") is None + assert await cp.aget_cache_value("ns", "k") is None + assert await cp.aclear_cache_value("ns", "k") is None + assert await cp.alist_cache_keys("ns") == [] diff --git a/tests/checkpointer/test_pg_checkpointer_extra.py b/tests/checkpointer/test_pg_checkpointer_extra.py index aa95370..c094cdb 100644 --- a/tests/checkpointer/test_pg_checkpointer_extra.py +++ b/tests/checkpointer/test_pg_checkpointer_extra.py @@ -1,5 +1,6 @@ import json -from unittest.mock import AsyncMock, MagicMock +from enum import Enum +from unittest.mock import AsyncMock, MagicMock, patch import pytest @@ -195,3 +196,162 @@ async def test_aget_message_not_found_raises(cp): with pytest.raises(ValueError): await cp.aget_message({"thread_id": "t1"}, "missing") + + +def test_import_error_asyncpg(monkeypatch): + monkeypatch.setattr("agentflow.storage.checkpointer.pg_checkpointer.HAS_ASYNCPG", False) + with pytest.raises(ImportError) as exc: + PgCheckpointer(postgres_dsn="postgres://x") + assert "requires 'asyncpg'" in str(exc.value) + + +def test_import_error_redis(monkeypatch): + monkeypatch.setattr("agentflow.storage.checkpointer.pg_checkpointer.HAS_ASYNCPG", True) + monkeypatch.setattr("agentflow.storage.checkpointer.pg_checkpointer.HAS_REDIS", False) + with pytest.raises(ImportError) as exc: + PgCheckpointer(postgres_dsn="postgres://x") + assert "requires 'redis'" in str(exc.value) + + +def test_schema_name_validation_on_init(monkeypatch): + monkeypatch.setattr("agentflow.storage.checkpointer.pg_checkpointer.HAS_ASYNCPG", True) + monkeypatch.setattr("agentflow.storage.checkpointer.pg_checkpointer.HAS_REDIS", True) + with pytest.raises(ValueError): + PgCheckpointer(postgres_dsn="postgres://x", redis=MagicMock(), schema="invalid-schema-name") + + +def test_init_with_pools(monkeypatch): + monkeypatch.setattr("agentflow.storage.checkpointer.pg_checkpointer.HAS_ASYNCPG", True) + monkeypatch.setattr("agentflow.storage.checkpointer.pg_checkpointer.HAS_REDIS", True) + + mock_pg_pool = MagicMock() + mock_redis_pool = MagicMock() + + with patch("agentflow.storage.checkpointer.pg_checkpointer.Redis") as mock_redis_class: + cp = PgCheckpointer(pg_pool=mock_pg_pool, redis_pool=mock_redis_pool) + assert cp._pg_pool is mock_pg_pool + mock_redis_class.assert_called_once_with(connection_pool=mock_redis_pool) + + +def test_create_redis_pool_no_url(monkeypatch): + monkeypatch.setattr("agentflow.storage.checkpointer.pg_checkpointer.HAS_ASYNCPG", True) + monkeypatch.setattr("agentflow.storage.checkpointer.pg_checkpointer.HAS_REDIS", True) + + cp = PgCheckpointer(postgres_dsn="postgres://x", redis=MagicMock()) + with pytest.raises(ValueError): + cp._create_redis_pool(redis=None, redis_pool=None, redis_url=None, redis_pool_config={}) + + +def test_create_pg_pool(cp): + mock_pool = MagicMock() + assert cp._create_pg_pool(pg_pool=mock_pool, postgres_dsn=None, pool_config={}) is mock_pool + + with patch("asyncpg.create_pool") as mock_create_pool: + cp._create_pg_pool(pg_pool=None, postgres_dsn="postgres://url", pool_config={"min_size": 5}) + mock_create_pool.assert_called_once_with(dsn="postgres://url", min_size=5) + + +@pytest.mark.asyncio +async def test_get_pg_pool_lazy(monkeypatch): + monkeypatch.setattr("agentflow.storage.checkpointer.pg_checkpointer.HAS_ASYNCPG", True) + monkeypatch.setattr("agentflow.storage.checkpointer.pg_checkpointer.HAS_REDIS", True) + + cp = PgCheckpointer(postgres_dsn="postgres://dsn", redis=MagicMock()) + assert cp._pg_pool is None + + mock_pool = MagicMock() + async def mock_create_pool(*args, **kwargs): + return mock_pool + + monkeypatch.setattr(cp, "_create_pg_pool", mock_create_pool) + pool = await cp._get_pg_pool() + assert pool is mock_pool + assert cp._pg_pool is mock_pool + + +def test_json_serializer_fast_json_importers(monkeypatch, cp): + monkeypatch.setenv("FAST_JSON", "1") + + import builtins + real_import = builtins.__import__ + def mock_import(name, *args, **kwargs): + if name in ("orjson", "msgspec"): + raise ImportError + return real_import(name, *args, **kwargs) + + with patch("builtins.__import__", side_effect=mock_import): + serializer = cp._get_json_serializer() + assert serializer == json.dumps + + +@pytest.mark.asyncio +async def test_check_and_apply_schema_version_upgrade(cp): + conn = AsyncMock() + conn.fetchrow = AsyncMock(return_value={"version": 1}) + + with patch.object(cp, "_get_current_schema_version", return_value=2): + await cp._check_and_apply_schema_version(conn) + + conn.execute.assert_called_once_with( + 'INSERT INTO "public"."schema_version" (version) VALUES ($1)', 2 + ) + + +@pytest.mark.asyncio +async def test_check_and_apply_schema_version_exception(cp): + conn = AsyncMock() + conn.fetchrow = AsyncMock(side_effect=RuntimeError("db error")) + + await cp._check_and_apply_schema_version(conn) + conn.execute.assert_called_once() + + +@pytest.mark.asyncio +async def test_initialize_schema_early_return(cp): + cp._schema_initialized = True + await cp._initialize_schema() + assert cp._pg_pool is None or not cp._pg_pool.acquire.called + + +@pytest.mark.asyncio +async def test_initialize_schema_error(cp): + cp._schema_initialized = False + + conn = AsyncMock() + conn.execute = AsyncMock(side_effect=RuntimeError("sql error")) + + cp._pg_pool = MagicMock() + cp._pg_pool.acquire.return_value = _AcquireCtx(conn) + + with pytest.raises(RuntimeError): + await cp._initialize_schema() + + +def test_serialize_state_enum_handler(cp): + class MockEnum(Enum): + VAL = "enum_val" + + class MockObj: + def __str__(self): + return "str_obj" + + mock_state = MagicMock() + mock_state.model_dump.return_value = {"enum": MockEnum.VAL, "obj": MockObj()} + serialized = cp._serialize_state(mock_state) + loaded = json.loads(serialized) + assert loaded["enum"] == "enum_val" + assert loaded["obj"] == "str_obj" + + +def test_deserialize_state_errors(cp): + class BadState: + @classmethod + def model_validate(cls, d): + raise TypeError("validation error") + + with pytest.raises(TypeError): + cp._deserialize_state({"invalid": "data"}, BadState) + + with pytest.raises(json.JSONDecodeError): + cp._deserialize_state("invalid-str", BadState) + diff --git a/tests/evaluation/test_phase3_evaluator.py b/tests/evaluation/test_phase3_evaluator.py index ea65c55..8e36e31 100644 --- a/tests/evaluation/test_phase3_evaluator.py +++ b/tests/evaluation/test_phase3_evaluator.py @@ -482,3 +482,305 @@ def test_render_error_case(self): html = reporter._render_case(result) assert "error" in html assert "Something went wrong" in html + + +# ============================================================================ +# Additional Reporter Tests for Coverage +# ============================================================================ + +from agentflow.qa.evaluation.dataset.eval_set import ToolCall, TrajectoryStep, StepType +from agentflow.qa.evaluation.eval_result import NodeDetail +from agentflow.qa.evaluation.token_usage import TokenUsage + +class NodeResponseObj: + node_name = "object_node" + response_text = "hello from object node" + tool_call_names = ["other_tool"] + is_final = False + has_tool_calls = False + timestamp = 140.0 + input_messages = [{"role": "system", "content": "system_prompt"}] + +def build_comprehensive_report(): + # 1. Tool Call + tc = ToolCall( + name="test_tool", + args={"arg1": "val1"}, + call_id="call_123", + result="success_result" + ) + + # 2. Trajectory Steps + step1 = TrajectoryStep( + step_type=StepType.TOOL, + name="test_tool", + args={"arg1": "val1"}, + timestamp=100.0, + metadata={"meta1": "val1"} + ) + + # 3. Node Response (dict format) + nr_dict = { + "node_name": "agent_node", + "response_text": "hello from node", + "tool_call_names": ["test_tool"], + "is_final": True, + "has_tool_calls": True, + "timestamp": 120.0, + "input_messages": [{"role": "user", "content": "hello"}] + } + + # 4. Node Detail (object format for node_details) + node_detail = NodeDetail( + node_name="other_node", + input_messages=[{"role": "user", "content": "hi"}], + response_text="hi response", + token_usage=TokenUsage(input_tokens=10, output_tokens=5), + timestamp=130.0 + ) + + # 5. Criterion Results + cr1 = CriterionResult( + criterion="traj_crit", + score=0.4, + passed=False, + threshold=0.8, + details={"reason": "traj failed reason", "extra": "extra_detail"}, + error="eval error message" + ) + cr2 = CriterionResult.success( + criterion="resp_crit", + score=0.9, + threshold=0.8, + details={"reason": "resp passed reason"} + ) + + # 6. Case Results + # Result 1: Failed + r1 = EvalCaseResult.success( + eval_id="case1", + name="Case One", + criterion_results=[cr1, cr2], + actual_trajectory=[step1], + actual_tool_calls=[tc], + actual_response="Hello world", + messages=[{"role": "user", "content": "query"}, {"role": "assistant", "content": "Hello world"}], + node_responses=[], + node_visits=["start_node", "agent_node"], + duration_seconds=2.5, + metadata={"case_meta": "meta_val"}, + turn_results=[{ + "turn_index": 0, + "user_input": "query", + "agent_response": "Hello world", + "tool_calls": [{"name": "test_tool"}], + "node_visits": ["start_node", "agent_node"] + }], + node_details=[node_detail] + ) + + # Assign attributes that bypass basic pydantic constructor validation + r1.actual_trajectory = [step1, "simple_trajectory_step"] + r1.node_responses = [nr_dict, NodeResponseObj()] + + # Result 2: Error + r2 = EvalCaseResult.failure( + eval_id="case2", + error="Case execution crash", + name="Case Two", + duration_seconds=1.2 + ) + + # 7. Create report + report = EvalReport.create( + eval_set_id="test_set_123", + eval_set_name="Comprehensive Test Set", + results=[r1, r2], + config_used={"eval_param": "value"} + ) + report.metadata = {"report_meta": "val"} + return report + +def test_colors_disable(): + # Save all original attributes from Colors + orig_attrs = {k: getattr(Colors, k) for k in ["RED", "RESET", "BOLD", "DIM", "GREEN", "YELLOW", "BLUE", "MAGENTA", "CYAN", "WHITE", "BG_RED", "BG_GREEN"]} + try: + Colors.disable() + assert Colors.RED == "" + finally: + for k, v in orig_attrs.items(): + setattr(Colors, k, v) + +def test_json_reporter_quick_save(): + report = build_comprehensive_report() + with tempfile.TemporaryDirectory() as tmpdir: + path = Path(tmpdir) / "quick.json" + JSONReporter.quick_save(report, str(path)) + assert path.exists() + +def test_json_reporter_generate(): + report = build_comprehensive_report() + reporter = JSONReporter() + + # 1. output_dir = None -> returns JSON string + json_str = reporter.generate(report) + assert "Comprehensive Test Set" in json_str + + # 2. output_dir is provided -> saves to file and returns path + with tempfile.TemporaryDirectory() as tmpdir: + res_path = reporter.generate(report, output_dir=tmpdir) + assert Path(res_path).exists() + assert res_path.endswith("report.json") + +def test_json_reporter_exclusions(): + report = build_comprehensive_report() + + # Disable everything + reporter = JSONReporter( + include_details=False, + include_trajectory=False, + include_node_responses=False, + include_actual_response=False, + include_tool_call_details=False, + ) + data = reporter.to_dict(report) + + for result in data["results"]: + assert "actual_trajectory" not in result + assert "actual_tool_calls" not in result + assert "node_responses" not in result + assert "node_details" not in result + assert "actual_response" not in result + for cr in result.get("criterion_results", []): + assert "details" not in cr + +def test_junit_reporter_generate(): + report = build_comprehensive_report() + reporter = JUnitXMLReporter() + + # 1. output_dir = None -> returns XML string + xml_str = reporter.generate(report) + assert "Comprehensive Test Set" in xml_str + + # 2. output_dir is provided -> saves to file and returns path + with tempfile.TemporaryDirectory() as tmpdir: + res_path = reporter.generate(report, output_dir=tmpdir) + assert Path(res_path).exists() + assert res_path.endswith("junit.xml") + +def test_junit_reporter_details(): + report = build_comprehensive_report() + reporter = JUnitXMLReporter() + xml_str = reporter.to_xml(report) + + assert "config_used" in xml_str + assert "report_meta" in xml_str + assert '' in xml_str + assert ' 127 for c in text): + raise UnicodeEncodeError("ascii", text, 0, len(text), "non-ascii") + return len(text) + + mock_output.write.side_effect = write_side_effect + + reporter = ConsoleReporter(use_color=False, output=mock_output) + reporter.report(report) + +def test_console_reporter_partial_stats(capsys): + cr_passed = CriterionResult.success( + criterion="yellow_crit", + score=0.9, + threshold=0.8, + ) + cr_failed = CriterionResult.success( + criterion="yellow_crit", + score=0.5, + threshold=0.8, + ) + r1 = EvalCaseResult.success( + eval_id="case1", + criterion_results=[cr_passed], + ) + r2 = EvalCaseResult.success( + eval_id="case2", + criterion_results=[cr_failed], + ) + + report = EvalReport.create( + eval_set_id="partial_set", + results=[r1, r2], + ) + report.summary.pass_rate = 0.5 + report.summary.criterion_stats = { + "yellow_crit": { + "pass_rate": 0.7, + "avg_score": 0.75, + "passed": 7, + "total": 10, + }, + "red_crit": { + "pass_rate": 0.3, + "avg_score": 0.35, + "passed": 3, + "total": 10, + } + } + + reporter = ConsoleReporter(use_color=False, verbose=True) + reporter.report(report) + + captured = capsys.readouterr() + assert "PARTIAL" in captured.out + assert "yellow_crit" in captured.out + assert "red_crit" in captured.out + +def test_console_reporter_comprehensive_printing(capsys): + report = build_comprehensive_report() + + reporter = ConsoleReporter( + use_color=True, + verbose=True, + include_trajectory=True, + include_actual_response=True, + ) + reporter.report(report) + + captured = capsys.readouterr() + assert "Comprehensive Test Set" in captured.out + assert "Case One" in captured.out + assert "Case Two" in captured.out + assert "ERROR" in captured.out + assert "Case execution crash" in captured.out + assert "case_meta" in captured.out + assert "test_tool" in captured.out + assert "call_id" in captured.out + assert "simple_trajectory_step" in captured.out + assert "object_node" in captured.out + assert "hello from object node" in captured.out + assert "Turn 0:" in captured.out + assert "traj failed reason" in captured.out + assert "extra_detail" in captured.out diff --git a/tests/evaluation/test_testing_fixtures.py b/tests/evaluation/test_testing_fixtures.py new file mode 100644 index 0000000..d1b7de9 --- /dev/null +++ b/tests/evaluation/test_testing_fixtures.py @@ -0,0 +1,182 @@ +"""Tests for Pytest integration utilities in testing.py.""" + +import pytest +from unittest.mock import AsyncMock, MagicMock, patch +from typing import Any +from pathlib import Path + +from agentflow.qa.evaluation.testing import ( + EvalTestCase, + eval_test, + assert_eval_passed, + assert_criterion_passed, + parametrize_eval_cases, + EvalFixtures, + EvalPlugin, + run_eval, + create_eval_app, + create_simple_eval_set, +) +from agentflow.qa.evaluation.eval_result import EvalReport +from agentflow.core.graph.compiled_graph import CompiledGraph + +def test_eval_test_case_repr(): + case = EvalTestCase(eval_id="id1", name="name1", description="desc1") + assert case.eval_id == "id1" + assert case.name == "name1" + assert case.description == "desc1" + assert repr(case) == "EvalTestCase(name1)" + + case2 = EvalTestCase(eval_id="id2") + assert repr(case2) == "EvalTestCase(id2)" + +@pytest.mark.asyncio +async def test_eval_test_decorator_success(): + mock_report = MagicMock() + mock_report.summary.pass_rate = 1.0 + + mock_evaluator = MagicMock() + mock_evaluator.evaluate = AsyncMock(return_value=mock_report) + + with patch("agentflow.qa.evaluation.AgentEvaluator", return_value=mock_evaluator) as mock_class: + with patch("agentflow.qa.evaluation.testing.Path.exists", return_value=True): + @eval_test(eval_file="dummy.json", threshold=0.8) + async def my_test(): + return "graph", "collector" + + await my_test() + mock_class.assert_called_once() + mock_evaluator.evaluate.assert_called_once_with("dummy.json", verbose=True) + +@pytest.mark.asyncio +async def test_eval_test_decorator_skips(): + import _pytest.outcomes + + @eval_test(eval_file="dummy.json") + async def my_skip_test(): + return None + + with pytest.raises(_pytest.outcomes.Skipped): + await my_skip_test() + +@pytest.mark.asyncio +async def test_eval_test_decorator_fails_invalid_return(): + import _pytest.outcomes + + @eval_test(eval_file="dummy.json") + async def my_fail_test(): + return "not-a-tuple" + + with pytest.raises(_pytest.outcomes.Failed): + await my_fail_test() + +@pytest.mark.asyncio +async def test_eval_test_decorator_fails_threshold_not_met(): + import _pytest.outcomes + + mock_report = MagicMock() + mock_report.summary.pass_rate = 0.5 + mock_report.failed_cases = [ + MagicMock(eval_id="case_1", name="Case One", error="failed", failed_criteria=[]) + ] + + mock_evaluator = MagicMock() + mock_evaluator.evaluate = AsyncMock(return_value=mock_report) + + with patch("agentflow.qa.evaluation.AgentEvaluator", return_value=mock_evaluator): + with patch("agentflow.qa.evaluation.testing.Path.exists", return_value=True): + @eval_test(eval_file="dummy.json", threshold=0.9) + async def my_fail_threshold_test(): + return "graph", "collector" + + with pytest.raises(_pytest.outcomes.Failed): + await my_fail_threshold_test() + +@pytest.mark.asyncio +async def test_eval_test_decorator_auto_detect_path(): + mock_report = MagicMock() + mock_report.summary.pass_rate = 1.0 + mock_evaluator = MagicMock() + mock_evaluator.evaluate = AsyncMock(return_value=mock_report) + + with patch("agentflow.qa.evaluation.AgentEvaluator", return_value=mock_evaluator): + with patch("agentflow.qa.evaluation.testing.Path.exists", return_value=True): + @eval_test() + async def test_my_custom_scenario(): + return "graph", "collector" + + await test_my_custom_scenario() + +def test_assert_eval_passed(): + report = MagicMock() + report.summary.pass_rate = 0.9 + report.failed_cases = [MagicMock(eval_id="c1", name="", failed_criteria=[])] + + assert_eval_passed(report, min_pass_rate=0.8) + + with pytest.raises(AssertionError, match="Evaluation pass rate 90.0% below threshold 95.0%"): + assert_eval_passed(report, min_pass_rate=0.95) + +def test_assert_criterion_passed(): + report = MagicMock() + report.summary.criterion_stats = { + "accuracy": {"avg_score": 0.85} + } + + with pytest.raises(AssertionError, match="Criterion 'safety' not found"): + assert_criterion_passed(report, "safety") + + assert_criterion_passed(report, "accuracy", min_score=0.8) + + with pytest.raises(AssertionError, match="Criterion 'accuracy' average score 0.85 below minimum 0.90"): + assert_criterion_passed(report, "accuracy", min_score=0.9) + +def test_parametrize_eval_cases(): + mock_set = MagicMock() + mock_case = MagicMock(eval_id="case1") + mock_set.eval_cases = [mock_case] + + with patch("agentflow.qa.evaluation.dataset.eval_set.EvalSet.from_file", return_value=mock_set) as mock_from_file: + decorator = parametrize_eval_cases("dummy_path.json") + assert decorator is not None + mock_from_file.assert_called_once_with("dummy_path.json") + +def test_eval_fixtures(): + fixtures = EvalFixtures(default_config="my_config") + assert fixtures.default_config == "my_config" + + with patch("agentflow.qa.evaluation.AgentEvaluator") as mock_eval: + factory = fixtures.evaluator_factory() + factory("graph", "collector") + mock_eval.assert_called_once_with("graph", "collector", config="my_config") + +def test_eval_plugin_noop(): + plugin = EvalPlugin() + plugin.pytest_configure(None) + plugin.pytest_collection_modifyitems(None, None) + +@pytest.mark.asyncio +async def test_run_eval_success(): + with patch("agentflow.qa.evaluation.AgentEvaluator") as mock_eval_class: + mock_eval = mock_eval_class.return_value + mock_eval.evaluate = AsyncMock(return_value="eval_report") + + res = await run_eval("graph", "collector", "path.json") + assert res == "eval_report" + mock_eval.evaluate.assert_called_once_with("path.json", verbose=False) + +def test_create_eval_app(): + mock_graph = MagicMock() + mock_graph.compile.return_value = "compiled_app" + + app, collector = create_eval_app(mock_graph) + assert app == "compiled_app" + assert collector is not None + mock_graph.compile.assert_called_once() + +def test_create_simple_eval_set(): + eval_set = create_simple_eval_set("my_set_id", [("query", "expected", "test_name")]) + assert eval_set.eval_set_id == "my_set_id" + assert eval_set.name == "my_set_id" + assert len(eval_set.eval_cases) == 1 + assert eval_set.eval_cases[0].name == "test_name" diff --git a/tests/evaluation/test_user_simulator.py b/tests/evaluation/test_user_simulator.py new file mode 100644 index 0000000..79bfe88 --- /dev/null +++ b/tests/evaluation/test_user_simulator.py @@ -0,0 +1,235 @@ +"""Tests for UserSimulator and BatchSimulator.""" + +import pytest +from unittest.mock import AsyncMock, MagicMock, patch +from typing import Any +import json +import uuid + +from agentflow.qa.evaluation.simulators.user_simulator import ( + UserSimulator, + BatchSimulator, + ConversationScenario, + SimulationResult +) +from agentflow.qa.evaluation.config.eval_config import UserSimulatorConfig +from agentflow.qa.evaluation.criteria.base import BaseCriterion, CriterionResult +from agentflow.core.graph.compiled_graph import CompiledGraph +from agentflow.core.state import Message + +class MockCriterion(BaseCriterion): + """Mock criterion for evaluation.""" + def __init__(self, name="mock_criterion"): + super().__init__() + self.name = name + + async def evaluate(self, execution_result: Any, eval_case: Any) -> CriterionResult: + return CriterionResult( + criterion=self.name, + score=0.9, + passed=True, + details={"notes": "good response"} + ) + +@pytest.mark.asyncio +async def test_user_simulator_init(): + # 1. Init with defaults + sim = UserSimulator() + assert sim.model == "gemini/gemini-2.5-flash" + assert sim.temperature == 0.7 + assert sim.max_turns == 10 + assert sim.api_style == "responses" + + # 2. Init with config + config = UserSimulatorConfig( + model="gpt-4o", + temperature=0.5, + max_invocations=5 + ) + # Mocking config object to have api_style + config.api_style = "chat" + sim_config = UserSimulator(config=config) + assert sim_config.model == "gpt-4o" + assert sim_config.temperature == 0.5 + assert sim_config.max_turns == 5 + assert sim_config.api_style == "chat" + +@pytest.mark.asyncio +async def test_user_simulator_run_success(): + # Prepare scenario + scenario = ConversationScenario( + scenario_id="test_scen", + description="A test scenario", + starting_prompt="Hello agent", + goals=["Say hello back"], + max_turns=3 + ) + + # Mock CompiledGraph + mock_graph = MagicMock(spec=CompiledGraph) + msg = Message.text_message("Hello back!", role="assistant") + mock_graph.ainvoke = AsyncMock(return_value={"messages": [msg]}) + + # Mock call_llm response for goal check + mock_result_json = json.dumps({"achieved": True, "reasoning": "Agent said hello back"}) + + criterion = MockCriterion("test_criterion") + sim = UserSimulator(criteria=[criterion]) + + with patch("agentflow.qa.evaluation.simulators.user_simulator.call_llm") as mock_call: + mock_call.return_value = (mock_result_json, 10, 20, 0) + + result = await sim.run(mock_graph, scenario) + + assert result.completed is True + assert "Say hello back" in result.goals_achieved + assert result.turns == 1 + assert result.criterion_scores["test_criterion"] == 0.9 + +@pytest.mark.asyncio +async def test_user_simulator_run_no_starting_prompt(): + scenario = ConversationScenario( + scenario_id="test_scen_no_start", + description="Test no start", + starting_prompt="", + goals=["Finish"], + max_turns=2 + ) + + mock_graph = MagicMock(spec=CompiledGraph) + msg = Message.text_message("Got it", role="assistant") + mock_graph.ainvoke = AsyncMock(return_value={"messages": [msg]}) + + sim = UserSimulator() + + with patch("agentflow.qa.evaluation.simulators.user_simulator.call_llm") as mock_call: + mock_call.side_effect = [ + ("Start prompt", 5, 5, 0), + (json.dumps({"achieved": True, "reasoning": "Done"}), 10, 10, 0) + ] + + result = await sim.run(mock_graph, scenario) + assert result.completed is True + assert result.conversation[0]["content"] == "Start prompt" + +@pytest.mark.asyncio +async def test_user_simulator_run_max_turns(): + scenario = ConversationScenario( + scenario_id="test_max_turns", + description="Test max turns limit", + starting_prompt="Hello", + goals=["Goal never met"], + max_turns=2 + ) + + mock_graph = MagicMock(spec=CompiledGraph) + msg = Message.text_message("Agent reply", role="assistant") + mock_graph.ainvoke = AsyncMock(return_value={"messages": [msg]}) + + sim = UserSimulator() + + with patch("agentflow.qa.evaluation.simulators.user_simulator.call_llm") as mock_call: + mock_call.side_effect = [ + (json.dumps({"achieved": False, "reasoning": "Not yet"}), 5, 5, 0), + ("User follow-up", 5, 5, 0), + (json.dumps({"achieved": False, "reasoning": "Still not yet"}), 5, 5, 0), + ("User second follow-up", 5, 5, 0), + ] + + result = await sim.run(mock_graph, scenario) + assert result.completed is False + assert result.turns == 2 + assert len(result.conversation) == 4 + +@pytest.mark.asyncio +async def test_user_simulator_graph_exception(): + scenario = ConversationScenario( + scenario_id="test_graph_err", + starting_prompt="Hello", + goals=["Done"] + ) + mock_graph = MagicMock(spec=CompiledGraph) + mock_graph.ainvoke = AsyncMock(side_effect=RuntimeError("Graph crash")) + + sim = UserSimulator() + result = await sim.run(mock_graph, scenario) + assert result.completed is False + assert "Graph crash" in result.error + +@pytest.mark.asyncio +async def test_user_simulator_check_goals_fallback(): + sim = UserSimulator() + conversation = [ + {"role": "user", "content": "hello"}, + {"role": "assistant", "content": "the magic word is banana"} + ] + + with patch("agentflow.qa.evaluation.simulators.user_simulator.call_llm", side_effect=Exception("LLM down")): + achieved, usage = await sim._check_goals( + all_goals=["banana", "apple"], + achieved=[], + conversation=conversation + ) + assert "banana" in achieved + assert "apple" not in achieved + +@pytest.mark.asyncio +async def test_user_simulator_general_exception(): + scenario = ConversationScenario(scenario_id="crash") + sim = UserSimulator() + mock_graph = MagicMock(spec=CompiledGraph) + mock_graph.ainvoke = AsyncMock(return_value={"messages": []}) + + # Cause crash inside try block by raising TypeError in _generate_initial_message + with patch.object(sim, "_generate_initial_message", side_effect=TypeError("Crashed inside")): + result = await sim.run(mock_graph, scenario) + assert result.completed is False + assert "Crashed inside" in result.error + +def test_extract_response_fallback_paths(): + sim = UserSimulator() + # 1. empty result + assert sim._extract_response({}) == "" + # 2. messages is empty + assert sim._extract_response({"messages": []}) == "" + # 3. assistant content block list + msg1 = type("Message", (), { + "role": "assistant", + "content": [ + type("Block", (), {"text": "Hello"}), + type("Block", (), {}) + ] + }) + assert sim._extract_response({"messages": [msg1]}) == "Hello" + + # 4. plain dict format + msg2 = {"role": "assistant", "content": "Dict text"} + assert sim._extract_response({"messages": [msg2]}) == "Dict text" + +@pytest.mark.asyncio +async def test_batch_simulator(): + sim = UserSimulator() + batch = BatchSimulator(simulator=sim, max_concurrency=2) + + mock_graph = MagicMock(spec=CompiledGraph) + msg = Message.text_message("Done", role="assistant") + mock_graph.ainvoke = AsyncMock(return_value={"messages": [msg]}) + + scenarios = [ + ConversationScenario(scenario_id="s1", starting_prompt="p1", goals=["g1"]), + ConversationScenario(scenario_id="s2", starting_prompt="p2", goals=["g2"]) + ] + + with patch("agentflow.qa.evaluation.simulators.user_simulator.call_llm") as mock_call: + mock_call.return_value = (json.dumps({"achieved": True, "reasoning": "ok"}), 1, 1, 0) + + results = await batch.run_batch(mock_graph, scenarios) + assert len(results) == 2 + assert results[0].scenario_id == "s1" + assert results[1].scenario_id == "s2" + + summary = batch.summary(results) + assert summary["total_scenarios"] == 2 + assert summary["completed"] == 2 + assert summary["completion_rate"] == 1.0 + assert summary["errors"] == 0 diff --git a/tests/graph/test_agent_internal.py b/tests/graph/test_agent_internal.py index 4793423..ca4e4e8 100644 --- a/tests/graph/test_agent_internal.py +++ b/tests/graph/test_agent_internal.py @@ -30,6 +30,18 @@ ) from agentflow.storage.store.store_schema import MemorySearchResult, MemoryType +from agentflow.core.graph.agent_internal.execution import ( + _extract_cache_creation_tokens, + _extract_cache_read_tokens, + _extract_finish_reason, + _extract_input_tokens, + _extract_output_tokens, + _extract_reasoning_tokens, + _extract_response_id, + _extract_response_model, + _extract_response_text, +) + # ───────────────────────────────────────────────────────────────────────────── # Shared helpers @@ -363,6 +375,210 @@ def test_returns_empty_string_for_empty_content(self): assert agent._extract_prompt(messages) == "" +# ═════════════════════════════════════════════════════════════════════════════ +# Token / response extraction helpers +# ═════════════════════════════════════════════════════════════════════════════ + + +class TestExtractInputTokens: + def test_no_usage_returns_zero(self): + assert _extract_input_tokens(SimpleNamespace()) == 0 + + def test_prompt_tokens(self): + resp = SimpleNamespace(usage=SimpleNamespace(prompt_tokens=42)) + assert _extract_input_tokens(resp) == 42 + + def test_input_tokens(self): + resp = SimpleNamespace(usage=SimpleNamespace(input_tokens=99)) + assert _extract_input_tokens(resp) == 99 + + def test_prompt_tokens_preferred_over_input_tokens(self): + resp = SimpleNamespace(usage=SimpleNamespace(prompt_tokens=10, input_tokens=20)) + assert _extract_input_tokens(resp) == 10 + + +class TestExtractOutputTokens: + def test_no_usage_returns_zero(self): + assert _extract_output_tokens(SimpleNamespace()) == 0 + + def test_completion_tokens(self): + resp = SimpleNamespace(usage=SimpleNamespace(completion_tokens=42)) + assert _extract_output_tokens(resp) == 42 + + def test_output_tokens(self): + resp = SimpleNamespace(usage=SimpleNamespace(output_tokens=99)) + assert _extract_output_tokens(resp) == 99 + + def test_completion_tokens_preferred_over_output_tokens(self): + resp = SimpleNamespace(usage=SimpleNamespace(completion_tokens=10, output_tokens=20)) + assert _extract_output_tokens(resp) == 10 + + +class TestExtractCacheReadTokens: + def test_no_usage_returns_zero(self): + assert _extract_cache_read_tokens(SimpleNamespace()) == 0 + + def test_anthropic_cache_read(self): + resp = SimpleNamespace(usage=SimpleNamespace(cache_read_input_tokens=100)) + assert _extract_cache_read_tokens(resp) == 100 + + def test_openai_chat_cached_tokens(self): + resp = SimpleNamespace( + usage=SimpleNamespace(prompt_tokens_details=SimpleNamespace(cached_tokens=50)), + ) + assert _extract_cache_read_tokens(resp) == 50 + + def test_openai_responses_cached_tokens(self): + resp = SimpleNamespace( + usage=SimpleNamespace(input_tokens_details=SimpleNamespace(cached_tokens=75)), + ) + assert _extract_cache_read_tokens(resp) == 75 + + def test_google_cached_content_token_count(self): + # Google responses have both usage (no cache fields) and usage_metadata + resp = SimpleNamespace( + usage=SimpleNamespace(prompt_tokens=5, completion_tokens=10), + usage_metadata=SimpleNamespace(cached_content_token_count=200), + ) + assert _extract_cache_read_tokens(resp) == 200 + + def test_no_cache_hits_returns_zero(self): + resp = SimpleNamespace(usage=SimpleNamespace(prompt_tokens=5, completion_tokens=10)) + assert _extract_cache_read_tokens(resp) == 0 + + def test_cache_value_zero_returns_zero(self): + resp = SimpleNamespace(usage=SimpleNamespace(cache_read_input_tokens=0)) + assert _extract_cache_read_tokens(resp) == 0 + + +class TestExtractCacheCreationTokens: + def test_no_usage_returns_zero(self): + assert _extract_cache_creation_tokens(SimpleNamespace()) == 0 + + def test_cache_creation_tokens(self): + resp = SimpleNamespace(usage=SimpleNamespace(cache_creation_input_tokens=50)) + assert _extract_cache_creation_tokens(resp) == 50 + + def test_cache_creation_none_returns_zero(self): + resp = SimpleNamespace(usage=SimpleNamespace(cache_creation_input_tokens=None)) + assert _extract_cache_creation_tokens(resp) == 0 + + def test_cache_creation_zero_returns_zero(self): + resp = SimpleNamespace(usage=SimpleNamespace(cache_creation_input_tokens=0)) + assert _extract_cache_creation_tokens(resp) == 0 + + +class TestExtractReasoningTokens: + def test_no_usage_returns_zero(self): + assert _extract_reasoning_tokens(SimpleNamespace()) == 0 + + def test_openai_completion_details(self): + resp = SimpleNamespace( + usage=SimpleNamespace(completion_tokens_details=SimpleNamespace(reasoning_tokens=42)), + ) + assert _extract_reasoning_tokens(resp) == 42 + + def test_openai_responses_output_details(self): + resp = SimpleNamespace( + usage=SimpleNamespace(output_tokens_details=SimpleNamespace(reasoning_tokens=77)), + ) + assert _extract_reasoning_tokens(resp) == 77 + + def test_google_thoughts_token_count(self): + # Google responses have both usage and usage_metadata + resp = SimpleNamespace( + usage=SimpleNamespace(completion_tokens=10), + usage_metadata=SimpleNamespace(thoughts_token_count=120), + ) + assert _extract_reasoning_tokens(resp) == 120 + + def test_no_reasoning_tokens_returns_zero(self): + resp = SimpleNamespace(usage=SimpleNamespace(completion_tokens=10)) + assert _extract_reasoning_tokens(resp) == 0 + + +class TestExtractFinishReason: + def test_openai_chat_choices_finish_reason(self): + resp = SimpleNamespace(choices=[SimpleNamespace(finish_reason="stop")]) + assert _extract_finish_reason(resp) == "stop" + + def test_openai_responses_status(self): + resp = SimpleNamespace(status="completed") + assert _extract_finish_reason(resp) == "completed" + + def test_status_in_progress_is_skipped(self): + resp = SimpleNamespace(status="in_progress") + assert _extract_finish_reason(resp) == "" + + def test_anthropic_stop_reason(self): + resp = SimpleNamespace(stop_reason="end_turn") + assert _extract_finish_reason(resp) == "end_turn" + + def test_google_candidates_finish_reason_with_name(self): + resp = SimpleNamespace( + candidates=[SimpleNamespace(finish_reason=SimpleNamespace(name="STOP"))], + ) + assert _extract_finish_reason(resp) == "STOP" + + def test_google_candidates_int_finish_reason(self): + resp = SimpleNamespace(candidates=[SimpleNamespace(finish_reason=1)]) + assert _extract_finish_reason(resp) == "1" + + def test_no_match_returns_empty(self): + assert _extract_finish_reason(SimpleNamespace()) == "" + + def test_empty_status_string_skipped(self): + resp = SimpleNamespace(status="") + assert _extract_finish_reason(resp) == "" + + +class TestExtractResponseId: + def test_with_id(self): + resp = SimpleNamespace(id="chatcmpl-abc123") + assert _extract_response_id(resp) == "chatcmpl-abc123" + + def test_no_id_returns_empty(self): + assert _extract_response_id(SimpleNamespace()) == "" + + +class TestExtractResponseModel: + def test_with_model(self): + resp = SimpleNamespace(model="gpt-4o") + assert _extract_response_model(resp) == "gpt-4o" + + def test_no_model_returns_empty(self): + assert _extract_response_model(SimpleNamespace()) == "" + + +class TestExtractResponseText: + def test_text_attribute(self): + resp = SimpleNamespace(text=" Hello world ") + assert _extract_response_text(resp) == "Hello world" + + def test_output_text_attribute(self): + resp = SimpleNamespace(output_text=" Hi there ") + assert _extract_response_text(resp) == "Hi there" + + def test_choices_content(self): + resp = SimpleNamespace(choices=[SimpleNamespace(message=SimpleNamespace(content="From choice"))]) + assert _extract_response_text(resp) == "From choice" + + def test_no_text_found_returns_empty(self): + assert _extract_response_text(SimpleNamespace()) == "" + + def test_text_preferred_over_choices(self): + resp = SimpleNamespace(text="Direct", choices=[SimpleNamespace(message=SimpleNamespace(content="Choice"))]) + assert _extract_response_text(resp) == "Direct" + + def test_output_text_preferred_over_choices(self): + resp = SimpleNamespace(output_text="Output", choices=[SimpleNamespace(message=SimpleNamespace(content="Choice"))]) + assert _extract_response_text(resp) == "Output" + + def test_choices_empty_content_returns_empty(self): + resp = SimpleNamespace(choices=[SimpleNamespace(message=SimpleNamespace(content=""))]) + assert _extract_response_text(resp) == "" + + # ═════════════════════════════════════════════════════════════════════════════ # AgentOpenAIMixin – _call_openai # ═════════════════════════════════════════════════════════════════════════════ @@ -1014,6 +1230,118 @@ async def test_unsupported_output_type_raises(self): await agent._call_google([{"role": "user", "content": "hi"}]) +# ═════════════════════════════════════════════════════════════════════════════ +# AgentExecutionMixin – _call_llm (provider routing) +# ═════════════════════════════════════════════════════════════════════════════ + + +@pytest.mark.asyncio +class TestCallLLM: + async def test_openai_chat_style(self): + agent = _make_openai_agent(api_style="chat") + agent._call_openai = AsyncMock(return_value="chat_response") + result = await agent._call_llm([{"role": "user", "content": "hi"}]) + assert result == "chat_response" + assert agent._effective_api_style == "chat" + agent._call_openai.assert_awaited_once() + + async def test_openai_responses_style(self): + agent = _make_openai_agent(api_style="responses", model="o4-mini") + agent._call_openai_responses = AsyncMock(return_value="resp_response") + result = await agent._call_llm([{"role": "user", "content": "hi"}]) + assert result == "resp_response" + assert agent._effective_api_style == "responses" + agent._call_openai_responses.assert_awaited_once() + + async def test_openai_responses_with_output_schema_falls_back_to_chat(self): + agent = _make_openai_agent(api_style="responses", model="o4-mini", output_schema={"type": "object", "properties": {"answer": {"type": "string"}}}) + agent._call_openai = AsyncMock(return_value="chat_response") + result = await agent._call_llm([{"role": "user", "content": "hi"}]) + assert result == "chat_response" + assert agent._effective_api_style == "chat" + agent._call_openai.assert_awaited_once() + + async def test_openai_responses_with_base_url_succeeds(self): + agent = _make_openai_agent(api_style="responses", model="o4-mini", base_url="http://localhost:8000/v1") + agent._call_openai_responses = AsyncMock(return_value="resp_response") + result = await agent._call_llm([{"role": "user", "content": "hi"}]) + assert result == "resp_response" + assert agent._effective_api_style == "responses" + agent._call_openai_responses.assert_awaited_once() + + async def test_openai_responses_with_base_url_fallback_on_error(self): + agent = _make_openai_agent(api_style="responses", model="o4-mini", base_url="http://localhost:8000/v1") + agent._call_openai_responses = AsyncMock(side_effect=Exception("not supported")) + agent._call_openai = AsyncMock(return_value="fallback_chat") + result = await agent._call_llm([{"role": "user", "content": "hi"}]) + assert result == "fallback_chat" + assert agent._effective_api_style == "chat" + agent._call_openai_responses.assert_awaited_once() + agent._call_openai.assert_awaited_once() + + async def test_openai_chat_with_reasoning_effort(self): + agent = _make_openai_agent(api_style="chat", reasoning_config={"effort": "high"}) + agent._call_openai = AsyncMock(return_value="response") + await agent._call_llm([{"role": "user", "content": "hi"}]) + call_kwargs = agent._call_openai.call_args[1] + assert call_kwargs.get("reasoning_effort") == "high" + + async def test_openai_chat_with_reasoning_config_and_base_url(self): + agent = _make_openai_agent(api_style="chat", reasoning_config={"effort": "medium"}, base_url="http://localhost:8000/v1") + agent._call_openai = AsyncMock(return_value="response") + await agent._call_llm([{"role": "user", "content": "hi"}]) + call_kwargs = agent._call_openai.call_args[1] + extra_body = call_kwargs.get("extra_body", {}) + assert "reasoning" in extra_body + assert extra_body["reasoning"] == {"effort": "medium"} + + async def test_google_provider(self): + agent = _make_google_agent() + agent._call_google = AsyncMock(return_value="google_response") + result = await agent._call_llm([{"role": "user", "content": "hi"}]) + assert result == "google_response" + agent._call_google.assert_awaited_once() + + async def test_unsupported_provider_raises(self): + agent = _make_openai_agent() + agent.provider = "unsupported" + with pytest.raises(ValueError, match="Unsupported provider"): + await agent._call_llm([{"role": "user", "content": "hi"}]) + + async def test_openai_responses_output_schema_with_reasoning_and_base_url(self): + """Cover lines 406-410: output_schema + reasoning_config + base_url.""" + agent = _make_openai_agent( + api_style="responses", + model="o4-mini", + output_schema={"type": "object", "properties": {"a": {"type": "string"}}}, + reasoning_config={"effort": "high"}, + base_url="http://localhost:8000/v1", + ) + agent._call_openai = AsyncMock(return_value="chat_result") + result = await agent._call_llm([{"role": "user", "content": "hi"}]) + assert result == "chat_result" + assert agent._effective_api_style == "chat" + call_kwargs = agent._call_openai.call_args[1] + assert call_kwargs.get("reasoning_effort") == "high" + extra_body = call_kwargs.get("extra_body", {}) + assert extra_body.get("reasoning") == {"effort": "high"} + + async def test_openai_responses_base_url_fallback_with_reasoning(self): + """Cover line 429: reasoning_effort in base_url fallback path.""" + agent = _make_openai_agent( + api_style="responses", + model="o4-mini", + reasoning_config={"effort": "low"}, + base_url="http://localhost:8000/v1", + ) + agent._call_openai_responses = AsyncMock(side_effect=Exception("fail")) + agent._call_openai = AsyncMock(return_value="fallback") + result = await agent._call_llm([{"role": "user", "content": "hi"}]) + assert result == "fallback" + call_kwargs = agent._call_openai.call_args[1] + assert call_kwargs.get("reasoning_effort") == "low" + + # ═════════════════════════════════════════════════════════════════════════════ # AgentExecutionMixin – _setup_tools # ═════════════════════════════════════════════════════════════════════════════ @@ -1187,6 +1515,413 @@ def my_tool(x: str) -> str: assert names_first.count("my_tool") == 1 assert names_second.count("my_tool") == 1 + async def test_named_node_not_found_raises(self): + """container.call_factory returning None should raise RuntimeError.""" + agent = _make_openai_agent(tool_node="MISSING") + agent._tool_node = None + agent.tool_node_name = "MISSING" + + container = MagicMock() + container.call_factory.return_value = None + + with pytest.raises(RuntimeError, match="ToolNode named 'MISSING' was not found"): + await agent._resolve_tools(container) + + async def test_named_node_not_tool_node_raises(self): + """Resolved node whose func is not a ToolNode should raise.""" + agent = _make_openai_agent(tool_node="NOT_TOOL") + agent._tool_node = None + agent.tool_node_name = "NOT_TOOL" + + fake_node = MagicMock() + fake_node.func = "not_a_tool_node" # not a ToolNode instance + + container = MagicMock() + container.call_factory.return_value = fake_node + + with pytest.raises(RuntimeError, match="not a ToolNode"): + await agent._resolve_tools(container) + + async def test_named_node_key_error_raises(self): + """Cover lines 719-720: container.call_factory raises KeyError.""" + from injectq.utils.exceptions import DependencyNotFoundError + + agent = _make_openai_agent(tool_node="MISSING") + agent._tool_node = None + agent.tool_node_name = "MISSING" + + container = MagicMock() + container.call_factory.side_effect = KeyError("get_node") + + with pytest.raises(RuntimeError, match="ToolNode named 'MISSING' was not found"): + await agent._resolve_tools(container) + + async def test_named_node_dependency_not_found_raises(self): + """Cover lines 719-720: container.call_factory raises DependencyNotFoundError.""" + from injectq.utils.exceptions import DependencyNotFoundError + + agent = _make_openai_agent(tool_node="MISSING") + agent._tool_node = None + agent.tool_node_name = "MISSING" + + container = MagicMock() + container.call_factory.side_effect = DependencyNotFoundError("get_node") + + with pytest.raises(RuntimeError, match="ToolNode named 'MISSING' was not found"): + await agent._resolve_tools(container) + + +# ═════════════════════════════════════════════════════════════════════════════ +# AgentExecutionMixin – _resolve_media_in_messages +# ═════════════════════════════════════════════════════════════════════════════ + + +@pytest.mark.asyncio +class TestResolveMediaInMessages: + async def test_no_media_store_returns_messages_unchanged(self): + agent = _make_openai_agent() + agent.media_store = None + messages = [{"role": "user", "content": "hello"}] + result = await agent._resolve_media_in_messages(messages) + assert result is messages + + async def test_non_list_content_skipped(self): + agent = _make_openai_agent() + agent.media_store = MagicMock() + messages = [{"role": "user", "content": "plain text"}] + result = await agent._resolve_media_in_messages(messages) + assert result is messages + + async def test_non_agentflow_url_not_touched(self): + agent = _make_openai_agent() + agent.media_store = MagicMock() + messages = [{"role": "user", "content": [{"type": "image_url", "image_url": {"url": "https://example.com/img.png"}}]}] + result = await agent._resolve_media_in_messages(messages) + assert result[0]["content"][0]["image_url"]["url"] == "https://example.com/img.png" + + async def test_openai_image_ref_resolved(self): + agent = _make_openai_agent() + agent.media_store = MagicMock() + agent.model = "gpt-4o" + agent.provider = "openai" + + resolved = {"type": "image_url", "image_url": {"url": "data:image/png;base64,abc123"}} + mock_resolver = AsyncMock() + mock_resolver.resolve_for_openai = AsyncMock(return_value=resolved) + + messages = [{"role": "user", "content": [{"type": "image_url", "image_url": {"url": "agentflow://media/img_1", "mime_type": "image/png"}}]}] + + with patch("agentflow.storage.media.resolver.MediaRefResolver", return_value=mock_resolver): + result = await agent._resolve_media_in_messages(messages) + + assert result[0]["content"][0] is resolved + mock_resolver.resolve_for_openai.assert_awaited_once() + + async def test_google_image_ref_with_inline_data(self): + agent = _make_openai_agent() + agent.media_store = MagicMock() + agent.model = "gemini-2.0-flash" + agent.provider = "google" + + resolved = SimpleNamespace( + inline_data=SimpleNamespace( + data=b"\x89PNG\r\n\x1a\n", + mime_type="image/png", + ), + ) + + mock_resolver = AsyncMock() + mock_resolver.resolve_for_google = AsyncMock(return_value=resolved) + + messages = [{"role": "user", "content": [{"type": "image_url", "image_url": {"url": "agentflow://media/img_g", "mime_type": "image/png"}}]}] + + with patch("agentflow.storage.media.resolver.MediaRefResolver", return_value=mock_resolver): + result = await agent._resolve_media_in_messages(messages) + + content = result[0]["content"][0] + assert content["type"] == "image_url" + assert content["image_url"]["url"].startswith("data:image/png;base64,") + + async def test_google_image_ref_with_file_data(self): + agent = _make_openai_agent() + agent.media_store = MagicMock() + agent.model = "gemini-2.0-flash" + agent.provider = "google" + + resolved = SimpleNamespace( + file_data=SimpleNamespace(file_uri="https://storage.googleapis.com/bucket/file"), + ) + + mock_resolver = AsyncMock() + mock_resolver.resolve_for_google = AsyncMock(return_value=resolved) + + messages = [{"role": "user", "content": [{"type": "image_url", "image_url": {"url": "agentflow://media/img_fd", "mime_type": "image/png"}}]}] + + with patch("agentflow.storage.media.resolver.MediaRefResolver", return_value=mock_resolver): + result = await agent._resolve_media_in_messages(messages) + + content = result[0]["content"][0] + assert content["type"] == "image_url" + assert content["image_url"]["url"] == "https://storage.googleapis.com/bucket/file" + + async def test_google_image_ref_resolve_failure_logged(self): + agent = _make_openai_agent() + agent.media_store = MagicMock() + agent.model = "gemini-2.0-flash" + agent.provider = "google" + + mock_resolver = AsyncMock() + mock_resolver.resolve_for_google = AsyncMock(side_effect=ValueError("API error")) + + messages = [{"role": "user", "content": [{"type": "image_url", "image_url": {"url": "agentflow://media/img_fail", "mime_type": "image/png"}}]}] + + with patch("agentflow.storage.media.resolver.MediaRefResolver", return_value=mock_resolver): + result = await agent._resolve_media_in_messages(messages) + + # Content unchanged on error + assert result[0]["content"][0]["image_url"]["url"] == "agentflow://media/img_fail" + + async def test_model_with_provider_prefix_stripped(self): + agent = _make_openai_agent() + agent.media_store = MagicMock() + agent.model = "openai/gpt-4o" + agent.provider = "openai" + + resolved = {"type": "image_url", "image_url": {"url": "data:...;base64,xyz"}} + mock_resolver = AsyncMock() + mock_resolver.resolve_for_openai = AsyncMock(return_value=resolved) + + messages = [{"role": "user", "content": [{"type": "image_url", "image_url": {"url": "agentflow://media/img_pfx", "mime_type": "image/png"}}]}] + + with patch("agentflow.storage.media.resolver.MediaRefResolver", return_value=mock_resolver): + await agent._resolve_media_in_messages(messages) + + # Model passed to resolver should not have the prefix + call_model = mock_resolver.resolve_for_openai.call_args[1].get("model") + assert call_model == "gpt-4o" + assert call_model != "openai/gpt-4o" + + async def test_document_type_openai_resolved(self): + agent = _make_openai_agent() + agent.media_store = MagicMock() + agent.model = "gpt-4o" + agent.provider = "openai" + + resolved = {"image_url": {"url": "data:application/pdf;base64,pdfdata"}} + mock_resolver = AsyncMock() + mock_resolver.resolve_for_openai = AsyncMock(return_value=resolved) + + messages = [{"role": "user", "content": [{"type": "document", "document": {"url": "agentflow://media/doc_1", "mime_type": "application/pdf"}}]}] + + with patch("agentflow.storage.media.resolver.MediaRefResolver", return_value=mock_resolver): + result = await agent._resolve_media_in_messages(messages) + + content = result[0]["content"][0] + assert content["type"] == "document" + assert content["document"]["url"] == "data:application/pdf;base64,pdfdata" + + async def test_video_type_google_resolved(self): + agent = _make_openai_agent() + agent.media_store = MagicMock() + agent.model = "gemini-2.0-flash" + agent.provider = "google" + + resolved = SimpleNamespace( + inline_data=SimpleNamespace( + data=b"videodata", + mime_type="video/mp4", + ), + ) + + mock_resolver = AsyncMock() + mock_resolver.resolve_for_google = AsyncMock(return_value=resolved) + + messages = [{"role": "user", "content": [{"type": "video", "video": {"url": "agentflow://media/vid_1", "mime_type": "video/mp4"}}]}] + + with patch("agentflow.storage.media.resolver.MediaRefResolver", return_value=mock_resolver): + result = await agent._resolve_media_in_messages(messages) + + content = result[0]["content"][0] + assert content["type"] == "video" + assert content["video"]["mime_type"] == "video/mp4" + + async def test_media_store_from_container_when_not_on_self(self): + agent = _make_openai_agent() + # Remove media_store from agent so it falls back to container lookup + if hasattr(agent, "media_store"): + del agent.media_store + + mock_media_store = MagicMock() + messages = [{"role": "user", "content": "hello"}] + + with patch("agentflow.core.graph.agent_internal.execution.InjectQ") as mock_injectq: + instance = mock_injectq.get_instance.return_value + instance.try_get.side_effect = lambda key: mock_media_store if key == "media_store" else None + result = await agent._resolve_media_in_messages(messages) + + assert result is messages # no agentflow refs, so unchanged + + async def test_non_dict_part_skipped(self): + """Cover line 595: part in content list is not a dict -> continue.""" + agent = _make_openai_agent() + agent.media_store = MagicMock() + messages = [{"role": "user", "content": ["not_a_dict", {"type": "image_url", "image_url": {"url": "https://example.com/img.png"}}]}] + result = await agent._resolve_media_in_messages(messages) + # Non-dict parts passed through unchanged + assert result[0]["content"][0] == "not_a_dict" + + async def test_document_type_openai_resolve_failure(self): + """Cover lines 696-697: exception during document/video resolve.""" + agent = _make_openai_agent() + agent.media_store = MagicMock() + agent.model = "gpt-4o" + agent.provider = "openai" + + mock_resolver = AsyncMock() + mock_resolver.resolve_for_openai = AsyncMock(side_effect=ValueError("doc resolve error")) + + messages = [{"role": "user", "content": [{"type": "document", "document": {"url": "agentflow://media/doc_fail", "mime_type": "application/pdf"}}]}] + + with patch("agentflow.storage.media.resolver.MediaRefResolver", return_value=mock_resolver): + result = await agent._resolve_media_in_messages(messages) + + assert result[0]["content"][0]["document"]["url"] == "agentflow://media/doc_fail" + + +# ═════════════════════════════════════════════════════════════════════════════ +# AgentExecutionMixin – execute +# ═════════════════════════════════════════════════════════════════════════════ + + +@pytest.mark.asyncio +class TestExecute: + async def test_basic_execute_text(self): + agent = _make_openai_agent() + state = AgentState() + config = {"_node_name": "AGENT", "is_stream": False} + + mock_response = SimpleNamespace( + id="resp_1", + model="gpt-4o", + text="Hello world", + usage=SimpleNamespace(prompt_tokens=10, completion_tokens=20), + ) + + agent._trim_context = AsyncMock(return_value=state) + agent._resolve_media_in_messages = AsyncMock(side_effect=lambda msgs: msgs) + agent._resolve_tools = AsyncMock(return_value=[]) + agent._call_llm_with_retry = AsyncMock(return_value=mock_response) + agent._build_skill_prompts = MagicMock(return_value=[]) + agent._build_memory_prompts = AsyncMock(return_value=[]) + + with patch("agentflow.core.graph.agent_internal.execution.convert_messages", return_value=[{"role": "user", "content": "hi"}]), \ + patch("agentflow.core.graph.agent_internal.execution.strip_media_blocks", side_effect=lambda msgs: msgs), \ + patch("agentflow.runtime.publisher.publish.publish_event") as mock_publish, \ + patch("agentflow.core.graph.agent_internal.execution.ModelResponseConverter") as mock_converter: + + mock_converter_instance = MagicMock() + mock_converter.return_value = mock_converter_instance + + result = await agent.execute(state, config) + + assert result is mock_converter_instance + agent._call_llm_with_retry.assert_awaited_once_with( + messages=[{"role": "user", "content": "hi"}], + tools=None, + stream=False, + ) + mock_converter.assert_called_once() + assert mock_converter.call_args.args[0] is mock_response + assert mock_converter.call_args.kwargs.get("converter") is not None + # START + END events published + assert mock_publish.call_count >= 2 + + async def test_execute_stream_no_end_event(self): + agent = _make_openai_agent() + state = AgentState() + config = {"_node_name": "AGENT", "is_stream": True} + + mock_response = MagicMock() + agent._trim_context = AsyncMock(return_value=state) + agent._resolve_media_in_messages = AsyncMock(side_effect=lambda msgs: msgs) + agent._resolve_tools = AsyncMock(return_value=[]) + agent._call_llm_with_retry = AsyncMock(return_value=mock_response) + agent._build_skill_prompts = MagicMock(return_value=[]) + agent._build_memory_prompts = AsyncMock(return_value=[]) + + with patch("agentflow.core.graph.agent_internal.execution.convert_messages", return_value=[]), \ + patch("agentflow.core.graph.agent_internal.execution.strip_media_blocks", side_effect=lambda msgs: msgs), \ + patch("agentflow.runtime.publisher.publish.publish_event") as mock_publish, \ + patch("agentflow.core.graph.agent_internal.execution.ModelResponseConverter") as mock_converter: + + await agent.execute(state, config) + + # Only START event — no END event for streams + assert mock_publish.call_count == 1 + + async def test_execute_with_multimodal_config_skips_strip(self): + agent = _make_openai_agent(multimodal_config=MagicMock()) + state = AgentState() + config = {"_node_name": "AGENT", "is_stream": False} + + mock_response = SimpleNamespace( + id="r1", model="gpt-4o", text="ok", + usage=SimpleNamespace(prompt_tokens=1, completion_tokens=1), + ) + + agent._trim_context = AsyncMock(return_value=state) + agent._resolve_media_in_messages = AsyncMock(side_effect=lambda msgs: msgs) + agent._resolve_tools = AsyncMock(return_value=[]) + agent._call_llm_with_retry = AsyncMock(return_value=mock_response) + + strip_called = [] + + def tracking_strip(msgs): + strip_called.append(True) + return msgs + + with patch("agentflow.core.graph.agent_internal.execution.convert_messages", return_value=[]), \ + patch("agentflow.core.graph.agent_internal.execution.strip_media_blocks", side_effect=tracking_strip), \ + patch("agentflow.runtime.publisher.publish.publish_event"), \ + patch("agentflow.core.graph.agent_internal.execution.ModelResponseConverter"): + + await agent.execute(state, config) + + # strip_media_blocks should NOT be called when multimodal_config is set + assert len(strip_called) == 0 + + async def test_execute_with_tools(self): + agent = _make_openai_agent() + state = AgentState() + config = {"_node_name": "AGENT", "is_stream": False} + + mock_response = SimpleNamespace( + id="r2", model="gpt-4o", text="tool result", + usage=SimpleNamespace(prompt_tokens=5, completion_tokens=15), + ) + + fake_tools = [{"function": {"name": "my_tool"}}] + + agent._trim_context = AsyncMock(return_value=state) + agent._resolve_media_in_messages = AsyncMock(side_effect=lambda msgs: msgs) + agent._resolve_tools = AsyncMock(return_value=fake_tools) + agent._call_llm_with_retry = AsyncMock(return_value=mock_response) + agent._build_skill_prompts = MagicMock(return_value=[]) + agent._build_memory_prompts = AsyncMock(return_value=[]) + + with patch("agentflow.core.graph.agent_internal.execution.convert_messages", return_value=[]), \ + patch("agentflow.core.graph.agent_internal.execution.strip_media_blocks", side_effect=lambda msgs: msgs), \ + patch("agentflow.runtime.publisher.publish.publish_event"), \ + patch("agentflow.core.graph.agent_internal.execution.ModelResponseConverter"): + + await agent.execute(state, config) + + agent._call_llm_with_retry.assert_awaited_once_with( + messages=[], + tools=fake_tools, + stream=False, + ) + # ═════════════════════════════════════════════════════════════════════════════ # Agent.__init__ – construction edge cases diff --git a/tests/graph/test_agent_retry_fallback.py b/tests/graph/test_agent_retry_fallback.py index 3040e02..f17ce1b 100644 --- a/tests/graph/test_agent_retry_fallback.py +++ b/tests/graph/test_agent_retry_fallback.py @@ -229,6 +229,23 @@ def test_string_fallback_429(self): exc = Exception("Rate limited: 429 Too Many Requests") assert agent._extract_status_code(exc) == 429 + def test_non_integer_code_returns_none(self): + agent = _make_agent() + exc = Exception("some error") + exc.code = "UNKNOWN" + assert agent._extract_status_code(exc) is None + + def test_none_code_returns_none(self): + agent = _make_agent() + exc = Exception("some error") + exc.code = None + assert agent._extract_status_code(exc) is None + + def test_string_with_no_code_returns_none(self): + agent = _make_agent() + exc = Exception("Something went wrong without a status code") + assert agent._extract_status_code(exc) is None + # ═════════════════════════════════════════════════════════════════════════════ # _is_retryable_error diff --git a/tests/graph/test_token_tracking.py b/tests/graph/test_token_tracking.py new file mode 100644 index 0000000..f8eef3e --- /dev/null +++ b/tests/graph/test_token_tracking.py @@ -0,0 +1,346 @@ +"""Tests for token tracking functionality in graph execution handlers. + +This module tests: +- Token usage calculation from agent state +- Token tracking in StreamHandler +- Token tracking in InvokeHandler +""" + +import pytest + +from agentflow.core.graph import StateGraph +from agentflow.core.graph.utils.utils import calculate_token_usage +from agentflow.core.state import AgentState, Message, TokenUsages +from agentflow.core.state.message_block import TextBlock +from agentflow.utils import END, ResponseGranularity + + +class TestCalculateTokenUsage: + """Test the calculate_token_usage utility function.""" + + def test_empty_context(self): + """Test token calculation with no messages.""" + state = AgentState() + usage = calculate_token_usage(state.context if state.context else []) + + assert usage["total_input_tokens"] == 0 # noqa: S101 + assert usage["total_output_tokens"] == 0 # noqa: S101 + assert usage["total_reasoning_tokens"] == 0 # noqa: S101 + assert usage["total_tokens"] == 0 # noqa: S101 + + def test_single_message_with_usage(self): + """Test token calculation with a single message.""" + message = Message( + role="assistant", + content=[TextBlock(text="Hello")], + usages=TokenUsages( + prompt_tokens=10, + completion_tokens=20, + total_tokens=30, + reasoning_tokens=5, + ), + ) + + state = AgentState(context=[message]) + usage = calculate_token_usage(state.context) + + assert usage["total_input_tokens"] == 10 # noqa: S101 + assert usage["total_output_tokens"] == 20 # noqa: S101 + assert usage["total_reasoning_tokens"] == 5 # noqa: S101 + assert usage["total_tokens"] == 30 # noqa: S101 + + def test_multiple_messages_with_usage(self): + """Test token calculation with multiple messages.""" + messages = [ + Message( + role="user", + content=[TextBlock(text="Hello")], + usages=TokenUsages( + prompt_tokens=5, + completion_tokens=0, + total_tokens=5, + reasoning_tokens=0, + ), + ), + Message( + role="assistant", + content=[TextBlock(text="Hi there!")], + usages=TokenUsages( + prompt_tokens=10, + completion_tokens=15, + total_tokens=25, + reasoning_tokens=3, + ), + ), + Message( + role="user", + content=[TextBlock(text="How are you?")], + usages=TokenUsages( + prompt_tokens=8, + completion_tokens=0, + total_tokens=8, + reasoning_tokens=0, + ), + ), + Message( + role="assistant", + content=[TextBlock(text="I'm doing well, thanks!")], + usages=TokenUsages( + prompt_tokens=12, + completion_tokens=20, + total_tokens=32, + reasoning_tokens=5, + ), + ), + ] + + state = AgentState(context=messages) + usage = calculate_token_usage(state.context) + + # Total: 5 + 10 + 8 + 12 = 35 input tokens + assert usage["total_input_tokens"] == 35 # noqa: S101 + # Total: 0 + 15 + 0 + 20 = 35 output tokens + assert usage["total_output_tokens"] == 35 # noqa: S101 + # Total: 0 + 3 + 0 + 5 = 8 reasoning tokens + assert usage["total_reasoning_tokens"] == 8 # noqa: S101 + # Total: 35 + 35 = 70 tokens + assert usage["total_tokens"] == 70 # noqa: S101 + + def test_messages_without_usage(self): + """Test token calculation with messages that don't have usage data.""" + messages = [ + Message( + role="user", + content=[TextBlock(text="Hello")], + usages=None, + ), + Message( + role="assistant", + content=[TextBlock(text="Hi there!")], + usages=TokenUsages( + prompt_tokens=10, + completion_tokens=15, + total_tokens=25, + reasoning_tokens=3, + ), + ), + ] + + state = AgentState(context=messages) + usage = calculate_token_usage(state.context) + + # Only the second message has usage + assert usage["total_input_tokens"] == 10 # noqa: S101 + assert usage["total_output_tokens"] == 15 # noqa: S101 + assert usage["total_reasoning_tokens"] == 3 # noqa: S101 + assert usage["total_tokens"] == 25 # noqa: S101 + + def test_mixed_messages_with_and_without_usage(self): + """Test token calculation with mix of messages with and without usage.""" + messages = [ + Message(role="user", content=[TextBlock(text="Hello")], usages=None), + Message( + role="assistant", + content=[TextBlock(text="Hi")], + usages=TokenUsages( + prompt_tokens=10, completion_tokens=5, total_tokens=15, reasoning_tokens=2 + ), + ), + Message(role="user", content=[TextBlock(text="Thanks")], usages=None), + Message( + role="assistant", + content=[TextBlock(text="You're welcome")], + usages=TokenUsages( + prompt_tokens=8, completion_tokens=12, total_tokens=20, reasoning_tokens=1 + ), + ), + ] + + state = AgentState(context=messages) + usage = calculate_token_usage(state.context) + + assert usage["total_input_tokens"] == 18 # noqa: S101 + assert usage["total_output_tokens"] == 17 # noqa: S101 + assert usage["total_reasoning_tokens"] == 3 # noqa: S101 + assert usage["total_tokens"] == 35 # noqa: S101 + + +@pytest.mark.asyncio +class TestStreamHandlerTokenTracking: + """Test token tracking in StreamHandler during graph execution.""" + + async def test_stream_handler_includes_token_usage_in_final_chunk(self): + """Test that StreamHandler includes token usage in final stream chunk.""" + from agentflow.core.state.stream_chunks import StreamEvent + + # Define a simple node that returns a message with token usage + def agent_node(state: AgentState, config: dict) -> list[Message]: + return [ + Message( + role="assistant", + content=[TextBlock(text="Hello from agent")], + usages=TokenUsages( + prompt_tokens=100, + completion_tokens=50, + total_tokens=150, + reasoning_tokens=10, + ), + ) + ] + + # Build and compile graph + graph = StateGraph(AgentState) + graph.add_node("agent", agent_node) + graph.set_entry_point("agent") + graph.add_edge("agent", END) + compiled_graph = graph.compile() + + # Execute with streaming + chunks = [] + async for chunk in compiled_graph.astream( + {"messages": [Message(role="user", content=[TextBlock(text="Hi")])]}, + response_granularity=ResponseGranularity.FULL, + ): + chunks.append(chunk) + + # Find the final UPDATES chunk + final_chunks = [c for c in chunks if c.event == StreamEvent.UPDATES and c.data.get("status") == "graph_invoked"] + assert len(final_chunks) > 0 # noqa: S101 + + final_chunk = final_chunks[-1] + + # Verify token usage is included + assert "total_input_tokens" in final_chunk.data # noqa: S101 + assert "total_output_tokens" in final_chunk.data # noqa: S101 + assert "total_reasoning_tokens" in final_chunk.data # noqa: S101 + assert "total_tokens" in final_chunk.data # noqa: S101 + + # Verify token counts (should match what we set in the message) + assert final_chunk.data["total_input_tokens"] == 100 # noqa: S101 + assert final_chunk.data["total_output_tokens"] == 50 # noqa: S101 + assert final_chunk.data["total_reasoning_tokens"] == 10 # noqa: S101 + assert final_chunk.data["total_tokens"] == 150 # noqa: S101 + + +@pytest.mark.asyncio +class TestInvokeHandlerTokenTracking: + """Test token tracking in InvokeHandler during graph execution.""" + + async def test_invoke_handler_includes_token_usage_in_response(self): + """Test that InvokeHandler includes token usage in response metadata.""" + + # Define a simple node that returns a message with token usage + def agent_node(state: AgentState, config: dict) -> list[Message]: + return [ + Message( + role="assistant", + content=[TextBlock(text="Response from agent")], + usages=TokenUsages( + prompt_tokens=200, + completion_tokens=100, + total_tokens=300, + reasoning_tokens=20, + ), + ) + ] + + # Build and compile graph + graph = StateGraph(AgentState) + graph.add_node("agent", agent_node) + graph.set_entry_point("agent") + graph.add_edge("agent", END) + compiled_graph = graph.compile() + + # Execute with invoke + result = await compiled_graph.ainvoke( + {"messages": [Message(role="user", content=[TextBlock(text="Hi")])]}, + response_granularity=ResponseGranularity.FULL, + ) + + # The result should contain the final state + assert "state" in result # noqa: S101 + assert "token_usage" in result # noqa: S101 + + # Verify token counts match what we set + token_usage = result["token_usage"] + assert token_usage["total_input_tokens"] == 200 # noqa: S101 + assert token_usage["total_output_tokens"] == 100 # noqa: S101 + assert token_usage["total_reasoning_tokens"] == 20 # noqa: S101 + assert token_usage["total_tokens"] == 300 # noqa: S101 # input + output only + + +@pytest.mark.asyncio +class TestMultipleNodesTokenTracking: + """Test token tracking across multiple nodes in a graph.""" + + async def test_token_accumulation_across_nodes(self): + """Test that tokens accumulate correctly across multiple nodes.""" + + def node1(state: AgentState, config: dict) -> list[Message]: + return [ + Message( + role="assistant", + content=[TextBlock(text="Node 1 response")], + usages=TokenUsages( + prompt_tokens=50, + completion_tokens=30, + total_tokens=80, + reasoning_tokens=5, + ), + ) + ] + + def node2(state: AgentState, config: dict) -> list[Message]: + return [ + Message( + role="assistant", + content=[TextBlock(text="Node 2 response")], + usages=TokenUsages( + prompt_tokens=60, + completion_tokens=40, + total_tokens=100, + reasoning_tokens=8, + ), + ) + ] + + def node3(state: AgentState, config: dict) -> list[Message]: + return [ + Message( + role="assistant", + content=[TextBlock(text="Node 3 response")], + usages=TokenUsages( + prompt_tokens=70, + completion_tokens=50, + total_tokens=120, + reasoning_tokens=10, + ), + ) + ] + + # Build graph with multiple nodes + graph = StateGraph(AgentState) + graph.add_node("node1", node1) + graph.add_node("node2", node2) + graph.add_node("node3", node3) + graph.set_entry_point("node1") + graph.add_edge("node1", "node2") + graph.add_edge("node2", "node3") + graph.add_edge("node3", END) + compiled_graph = graph.compile() + + # Execute + result = await compiled_graph.ainvoke( + {"messages": [Message(role="user", content=[TextBlock(text="Start")])]}, + response_granularity=ResponseGranularity.FULL, + ) + + # Verify token usage is included in the result + assert "token_usage" in result # noqa: S101 + token_usage = result["token_usage"] + + # Verify accumulated tokens (50 + 60 + 70 = 180 input, 30 + 40 + 50 = 120 output, 5 + 8 + 10 = 23 reasoning) + assert token_usage["total_input_tokens"] == 180 # noqa: S101 + assert token_usage["total_output_tokens"] == 120 # noqa: S101 + assert token_usage["total_reasoning_tokens"] == 23 # noqa: S101 + assert token_usage["total_tokens"] == 300 # noqa: S101 # input + output only diff --git a/tests/publisher/test_composite_publisher.py b/tests/publisher/test_composite_publisher.py new file mode 100644 index 0000000..3a24107 --- /dev/null +++ b/tests/publisher/test_composite_publisher.py @@ -0,0 +1,43 @@ +import pytest +from unittest.mock import AsyncMock, MagicMock +from agentflow.runtime.publisher.composite_publisher import CompositePublisher +from agentflow.runtime.publisher.base_publisher import BasePublisher +from agentflow.runtime.publisher.events import EventModel + + +@pytest.mark.asyncio +async def test_composite_publisher_lifecycle(): + pub1 = AsyncMock(spec=BasePublisher) + pub2 = AsyncMock(spec=BasePublisher) + + # Initialize + composite = CompositePublisher([pub1]) + assert pub1 in composite._publishers + assert pub2 not in composite._publishers + + # Add publisher + composite.add_publisher(pub2) + assert pub2 in composite._publishers + + # Publish + event = MagicMock(spec=EventModel) + await composite.publish(event) + pub1.publish.assert_called_once_with(event) + pub2.publish.assert_called_once_with(event) + + # Close + await composite.close() + pub1.close.assert_called_once() + pub2.close.assert_called_once() + + # Sync close + pub1.sync_close = MagicMock() + pub2.sync_close = MagicMock() + composite.sync_close() + pub1.sync_close.assert_called_once() + pub2.sync_close.assert_called_once() + + # Remove publisher + composite.remove_publisher(pub1) + assert pub1 not in composite._publishers + assert pub2 in composite._publishers diff --git a/tests/publisher/test_optional_publishers.py b/tests/publisher/test_optional_publishers.py index 05e9592..3f20b14 100644 --- a/tests/publisher/test_optional_publishers.py +++ b/tests/publisher/test_optional_publishers.py @@ -411,6 +411,84 @@ async def test_kafka_publisher_publish_success(self, mock_import): assert parsed_message["event"] == "tool_execution" assert parsed_message["node_name"] == "kafka_test" + @pytest.mark.asyncio + async def test_kafka_publisher_closed_errors(self): + """Test KafkaPublisher closed errors.""" + from agentflow.runtime.publisher.kafka_publisher import KafkaPublisher + publisher = KafkaPublisher() + publisher._is_closed = True + + with pytest.raises(RuntimeError, match="KafkaPublisher is closed"): + await publisher._get_producer() + + with pytest.raises(RuntimeError, match="Cannot publish to closed KafkaPublisher"): + await publisher.publish(EventModel(event=Event.GRAPH_EXECUTION, event_type=EventType.START)) + + @pytest.mark.asyncio + @patch('importlib.import_module') + async def test_kafka_publisher_get_producer_early_return(self, mock_import): + """Test early return in _get_producer if already initialized.""" + from agentflow.runtime.publisher.kafka_publisher import KafkaPublisher + publisher = KafkaPublisher() + publisher._producer = MagicMock() + + res = await publisher._get_producer() + assert res == publisher._producer + mock_import.assert_not_called() + + @pytest.mark.asyncio + async def test_kafka_publisher_missing_module(self): + """Test missing aiokafka ImportError fallback.""" + from agentflow.runtime.publisher.kafka_publisher import KafkaPublisher + publisher = KafkaPublisher() + + with patch('agentflow.runtime.publisher.kafka_publisher.importlib.import_module', side_effect=ImportError): + with pytest.raises(RuntimeError, match="requires the 'aiokafka' package"): + await publisher._get_producer() + + @pytest.mark.asyncio + async def test_kafka_publisher_close_methods(self): + """Test close() and sync_close() paths.""" + from agentflow.runtime.publisher.kafka_publisher import KafkaPublisher + publisher = KafkaPublisher() + + # Closed early return + publisher._is_closed = True + await publisher.close() + + # Close with producer + publisher._is_closed = False + mock_producer = AsyncMock() + publisher._producer = mock_producer + + await publisher.close() + mock_producer.stop.assert_called_once() + assert publisher._producer is None + assert publisher._is_closed is True + + # Close with exception in stop + publisher = KafkaPublisher() + mock_producer = AsyncMock() + mock_producer.stop.side_effect = Exception("stop failed") + publisher._producer = mock_producer + + await publisher.close() + assert publisher._producer is None + assert publisher._is_closed is True + + # Test sync_close + publisher = KafkaPublisher() + publisher.close = AsyncMock() + publisher.sync_close() + publisher.close.assert_called_once() + + # Test sync_close active loop warning + publisher = KafkaPublisher() + with patch('asyncio.run', side_effect=RuntimeError): + with patch('agentflow.runtime.publisher.kafka_publisher.logger.warning') as mock_warn: + publisher.sync_close() + mock_warn.assert_called_once() + class TestRabbitMQPublisher: """Test RabbitMQPublisher with mocked dependencies.""" @@ -504,6 +582,124 @@ async def test_rabbitmq_publisher_publish_success(self, mock_import): # Verify message creation - Mock should be called as a constructor assert mock_pika.Message.called + + @pytest.mark.asyncio + async def test_rabbitmq_publisher_closed_errors(self): + """Test RabbitMQPublisher raising RuntimeError when closed.""" + from agentflow.runtime.publisher.rabbitmq_publisher import RabbitMQPublisher + publisher = RabbitMQPublisher() + publisher._is_closed = True + + with pytest.raises(RuntimeError, match="RabbitMQPublisher is closed"): + await publisher._ensure() + + with pytest.raises(RuntimeError, match="Cannot publish to closed RabbitMQPublisher"): + await publisher.publish(EventModel(event=Event.GRAPH_EXECUTION, event_type=EventType.START)) + + @pytest.mark.asyncio + @patch('importlib.import_module') + async def test_rabbitmq_publisher_ensure_early_return(self, mock_import): + """Test early return in _ensure when exchange is already declared.""" + from agentflow.runtime.publisher.rabbitmq_publisher import RabbitMQPublisher + publisher = RabbitMQPublisher() + publisher._exchange = MagicMock() + + await publisher._ensure() + mock_import.assert_not_called() + + @pytest.mark.asyncio + async def test_rabbitmq_publisher_missing_module(self): + """Test RuntimeError when aio_pika module is missing.""" + from agentflow.runtime.publisher.rabbitmq_publisher import RabbitMQPublisher + publisher = RabbitMQPublisher() + + with patch('agentflow.runtime.publisher.rabbitmq_publisher.importlib.import_module', side_effect=ImportError): + with pytest.raises(RuntimeError, match="requires the 'aio-pika' package"): + await publisher._ensure() + + @pytest.mark.asyncio + @patch('importlib.import_module') + async def test_rabbitmq_publisher_default_exchange_fallback(self, mock_import): + """Test default exchange declaration fallback when declare=False.""" + mock_pika = Mock() + mock_connection = AsyncMock() + mock_channel = AsyncMock() + + mock_pika.connect_robust = AsyncMock(return_value=mock_connection) + mock_connection.channel = AsyncMock(return_value=mock_channel) + mock_import.return_value = mock_pika + + with patch.dict('sys.modules', {'aio_pika': mock_pika}): + from agentflow.runtime.publisher.rabbitmq_publisher import RabbitMQPublisher + publisher = RabbitMQPublisher({"declare": False}) + + await publisher._ensure() + assert publisher._exchange == mock_channel.default_exchange + + @pytest.mark.asyncio + @patch('importlib.import_module') + async def test_rabbitmq_publisher_publish_not_initialized(self, mock_import): + """Test publish raises RuntimeError when exchange is not initialized.""" + mock_pika = Mock() + mock_import.return_value = mock_pika + from agentflow.runtime.publisher.rabbitmq_publisher import RabbitMQPublisher + publisher = RabbitMQPublisher() + + # Mock _ensure to do nothing so _exchange remains None + with patch.object(publisher, '_ensure', AsyncMock()): + with pytest.raises(RuntimeError, match="exchange not initialized"): + await publisher.publish(EventModel(event=Event.GRAPH_EXECUTION, event_type=EventType.START)) + + @pytest.mark.asyncio + async def test_rabbitmq_publisher_close_methods(self): + """Test close() and sync_close() paths.""" + from agentflow.runtime.publisher.rabbitmq_publisher import RabbitMQPublisher + publisher = RabbitMQPublisher() + + # Close when already closed + publisher._is_closed = True + await publisher.close() # Should return early + + # Close with active channel and connection + publisher._is_closed = False + mock_channel = AsyncMock() + mock_conn = AsyncMock() + publisher._channel = mock_channel + publisher._conn = mock_conn + + await publisher.close() + mock_channel.close.assert_called_once() + mock_conn.close.assert_called_once() + assert publisher._channel is None + assert publisher._conn is None + assert publisher._is_closed is True + + # Test close with exceptions on close call (should catch and log) + publisher = RabbitMQPublisher() + publisher._channel = AsyncMock() + publisher._channel.close.side_effect = Exception("channel close failed") + publisher._conn = AsyncMock() + publisher._conn.close.side_effect = Exception("conn close failed") + + await publisher.close() + assert publisher._channel is None + assert publisher._conn is None + assert publisher._is_closed is True + + # Test sync_close + publisher = RabbitMQPublisher() + publisher.close = AsyncMock() + publisher.sync_close() + publisher.close.assert_called_once() + + # Test sync_close raises RuntimeError (active loop) + publisher = RabbitMQPublisher() + with patch('asyncio.run', side_effect=RuntimeError("active loop")): + with patch('agentflow.runtime.publisher.rabbitmq_publisher.logger.warning') as mock_warn: + publisher.sync_close() + mock_warn.assert_called_once_with("sync_close called within an active event loop; skipping.") + + class TestOptionalPublisherErrorHandling: """Test error handling across optional publishers.""" diff --git a/tests/publisher/test_publish.py b/tests/publisher/test_publish.py new file mode 100644 index 0000000..29d4f12 --- /dev/null +++ b/tests/publisher/test_publish.py @@ -0,0 +1,42 @@ +import pytest +from unittest.mock import AsyncMock, MagicMock +from agentflow.runtime.publisher.publish import _publish_event_task, publish_event +from agentflow.runtime.publisher.events import EventModel +from agentflow.runtime.publisher.base_publisher import BasePublisher +from agentflow.utils.background_task_manager import BackgroundTaskManager + + +@pytest.mark.asyncio +async def test_publish_event_task_success(): + event = MagicMock(spec=EventModel) + publisher = AsyncMock(spec=BasePublisher) + + await _publish_event_task(event, publisher) + + publisher.publish.assert_called_once_with(event) + + +@pytest.mark.asyncio +async def test_publish_event_task_failure(): + event = MagicMock(spec=EventModel) + publisher = AsyncMock(spec=BasePublisher) + publisher.publish.side_effect = RuntimeError("publish boom") + + await _publish_event_task(event, publisher) + + publisher.publish.assert_called_once_with(event) + + +@pytest.mark.asyncio +async def test_publish_event_task_no_publisher(): + event = MagicMock(spec=EventModel) + await _publish_event_task(event, None) + + +def test_publish_event(): + event = MagicMock(spec=EventModel) + publisher = MagicMock(spec=BasePublisher) + task_manager = MagicMock(spec=BackgroundTaskManager) + + publish_event(event, publisher, task_manager) + task_manager.create_task.assert_called_once() diff --git a/tests/storage/media/test_base_media_store.py b/tests/storage/media/test_base_media_store.py new file mode 100644 index 0000000..fb7ba8a --- /dev/null +++ b/tests/storage/media/test_base_media_store.py @@ -0,0 +1,58 @@ +import pytest +from typing import Any +from agentflow.storage.media.storage.base import BaseMediaStore +from agentflow.core.state.message_block import MediaRef + +class ConcreteMediaStore(BaseMediaStore): + """A concrete implementation of BaseMediaStore for testing.""" + def __init__(self): + self.store_dict = {} + + async def store(self, data: bytes, mime_type: str, metadata: dict[str, Any] | None = None) -> str: + key = f"key_{len(self.store_dict)}" + self.store_dict[key] = (data, mime_type) + return key + + async def retrieve(self, storage_key: str) -> tuple[bytes, str]: + if storage_key not in self.store_dict: + raise KeyError(f"Key {storage_key} not found") + return self.store_dict[storage_key] + + async def delete(self, storage_key: str) -> bool: + if storage_key in self.store_dict: + del self.store_dict[storage_key] + return True + return False + + async def exists(self, storage_key: str) -> bool: + return storage_key in self.store_dict + +@pytest.mark.asyncio +async def test_get_metadata_success(): + store = ConcreteMediaStore() + key = await store.store(b"hello", "text/plain") + metadata = await store.get_metadata(key) + assert metadata == { + "mime_type": "text/plain", + "size_bytes": 5 + } + +@pytest.mark.asyncio +async def test_get_metadata_key_error(): + store = ConcreteMediaStore() + metadata = await store.get_metadata("nonexistent") + assert metadata is None + +@pytest.mark.asyncio +async def test_get_direct_url_default(): + store = ConcreteMediaStore() + url = await store.get_direct_url("key") + assert url is None + +def test_to_media_ref(): + store = ConcreteMediaStore() + ref = store.to_media_ref("key", "image/png") + assert isinstance(ref, MediaRef) + assert ref.kind == "url" + assert ref.url == "agentflow://media/key" + assert ref.mime_type == "image/png" diff --git a/tests/storage/media/test_media_resolver_extra.py b/tests/storage/media/test_media_resolver_extra.py index 14bff5f..85b2e5f 100644 --- a/tests/storage/media/test_media_resolver_extra.py +++ b/tests/storage/media/test_media_resolver_extra.py @@ -284,3 +284,132 @@ async def test_retrieve_bytes_invalid_kind_raises(): def test_openai_image_url_helper_shape(): part = _openai_image_url("https://example.com/x.png") assert part == {"type": "image_url", "image_url": {"url": "https://example.com/x.png"}} + + +def test_with_cache_configures_resolver(): + from agentflow.storage.media.resolver import MediaRefResolver + resolver = MediaRefResolver() + cache = object() + out = resolver.with_cache(cache, expiration_seconds=1800, refresh_buffer_seconds=30) + assert out is resolver + assert resolver.cache_backend is cache + assert resolver.direct_url_expiration_seconds == 1800 + assert resolver.direct_url_refresh_buffer_seconds == 30 + + +@pytest.mark.asyncio +async def test_resolve_openai_legacy_fallback_reftypes_and_empty(): + from agentflow.storage.media.resolver import MediaRefResolver + resolver = MediaRefResolver() + ref_empty = MediaRef.model_construct(kind="unknown") + res = await resolver._resolve_openai_legacy(ref_empty) + assert res == {"type": "image_url", "image_url": {"url": ""}} + + +@pytest.mark.asyncio +async def test_resolve_google_legacy_various_refs(): + from agentflow.storage.media.resolver import MediaRefResolver + resolver = MediaRefResolver(media_store=_Store()) + + class _Part: + @staticmethod + def from_uri(file_uri, mime_type): + return {"uri": file_uri, "mime": mime_type} + + @staticmethod + def from_bytes(data, mime_type): + return {"data": data, "mime": mime_type} + + def __init__(self, **kwargs): + self.kwargs = kwargs + + with patch("google.genai.types.Part", _Part): + ref_internal = MediaRef(kind="url", url="agentflow://media/k") + part1 = await resolver._resolve_google_legacy(ref_internal) + assert part1 == {"uri": "https://signed.example/k.png", "mime": "application/octet-stream"} + + resolver.media_store.url_map["k"] = None + part2 = await resolver._resolve_google_legacy(ref_internal) + assert part2 == {"data": b"abc", "mime": "image/png"} + + ref_external = MediaRef(kind="url", url="https://example.com/y.jpg", mime_type="image/jpeg") + part3 = await resolver._resolve_google_legacy(ref_external) + assert part3 == {"uri": "https://example.com/y.jpg", "mime": "image/jpeg"} + + ref_data = MediaRef(kind="data", data_base64=base64.b64encode(b"hello").decode(), mime_type="text/plain") + part4 = await resolver._resolve_google_legacy(ref_data) + assert part4 == {"data": b"hello", "mime": "text/plain"} + + ref_file = MediaRef(kind="file_id", file_id="file-123", mime_type="image/png") + part5 = await resolver._resolve_google_legacy(ref_file) + assert isinstance(part5, _Part) + assert part5.kwargs["file_data"].file_uri == "file-123" + + ref_unres = MediaRef.model_construct(kind="unknown") + part6 = await resolver._resolve_google_legacy(ref_unres) + assert isinstance(part6, _Part) + assert part6.kwargs["text"] == "[Unresolvable media reference]" + + +@pytest.mark.asyncio +async def test_try_transport_modes(): + from agentflow.storage.media.resolver import MediaRefResolver + resolver = MediaRefResolver() + + result = await resolver._try_transport(MediaRef(kind="url"), MediaTransportMode.provider_file, "openai", object()) + assert result is None + + class _Part: + @staticmethod + def from_uri(file_uri, mime_type): + return {"uri": file_uri, "mime": mime_type} + + resolver.media_store = _Store() + caps = type("Caps", (), {"can_convert_internal_to_remote": True})() + with patch("google.genai.types.Part", _Part): + res = await resolver._transport_remote_url( + MediaRef(kind="url", url="agentflow://media/k", mime_type="image/png"), + caps, + provider="google" + ) + assert res == {"uri": "https://signed.example/k.png", "mime": "image/png"} + + +@pytest.mark.asyncio +async def test_transport_inline_bytes_url_retrieve_fail(): + from agentflow.storage.media.resolver import MediaRefResolver + resolver = MediaRefResolver() + async def _fail(*args): + raise ValueError("fetch fail") + resolver._retrieve_bytes = _fail + res = await resolver._transport_inline_bytes(MediaRef(kind="url", url="https://x"), "openai") + assert res is None + + +@pytest.mark.asyncio +async def test_get_cached_signed_url_expired_or_invalid(): + from agentflow.storage.media.resolver import MediaRefResolver + resolver = MediaRefResolver(cache_backend=_Cache()) + resolver.cache_backend.values[("media:signed-url", "k")] = "not-a-dict" + res1 = await resolver._get_cached_signed_url("k") + assert res1 is None + + resolver.cache_backend.values[("media:signed-url", "k")] = {"url": "https://x"} + res2 = await resolver._get_cached_signed_url("k") + assert res2 is None + + resolver.cache_backend.values[("media:signed-url", "k")] = {"url": "https://x", "expires_at": 100} + resolver.direct_url_refresh_buffer_seconds = 60 + res3 = await resolver._get_cached_signed_url("k") + assert res3 is None + + +def test_source_kind_helper_variations(): + from agentflow.storage.media.resolver import _source_kind + assert _source_kind(MediaRef(kind="url", url="agentflow://media/k")) == "internal_ref" + assert _source_kind(MediaRef(kind="url", url="https://x")) == "url" + assert _source_kind(MediaRef(kind="data")) == "data" + assert _source_kind(MediaRef(kind="file_id")) == "file_id" + assert _source_kind(MediaRef.model_construct(kind="other")) == "other" + + diff --git a/tests/storage/media/test_provider_media.py b/tests/storage/media/test_provider_media.py new file mode 100644 index 0000000..7a03614 --- /dev/null +++ b/tests/storage/media/test_provider_media.py @@ -0,0 +1,108 @@ +import pytest +from unittest.mock import MagicMock, AsyncMock, patch +from agentflow.storage.media.provider_media import ( + ProviderMediaCache, + should_use_google_file_api, + prepare_google_content_part, + upload_to_google_file_api, + create_openai_file_search_tool, + create_openai_file_attachment +) + +def test_provider_media_cache(): + cache = ProviderMediaCache(max_entries=2) + key1 = cache.content_key(b"data1") + key2 = cache.content_key(b"data2") + key3 = cache.content_key(b"data3") + + cache.put("google", key1, "ref1") + cache.put("google", key2, "ref2") + assert cache.get("google", key1) == "ref1" + assert cache.get("google", key2) == "ref2" + + # Test eviction: key1 (oldest) should be evicted when key3 is added + cache.put("google", key3, "ref3") + assert cache.get("google", key1) is None + assert cache.get("google", key2) == "ref2" + assert cache.get("google", key3) == "ref3" + + # Test clear for specific provider + cache.clear("google") + assert cache.get("google", key2) is None + + # Test clear for all + cache.put("google", key2, "ref2") + cache.clear() + assert cache.get("google", key2) is None + +def test_should_use_google_file_api(): + # threshold is 20MB + assert should_use_google_file_api(10 * 1024 * 1024) is False + assert should_use_google_file_api(25 * 1024 * 1024) is True + +def test_prepare_google_content_part(): + class _Part: + @staticmethod + def from_bytes(data, mime_type): + return {"data": data, "mime": mime_type} + + with patch("google.genai.types.Part", _Part): + # Under threshold + res = prepare_google_content_part(b"abc", "image/png") + assert res == {"data": b"abc", "mime": "image/png"} + + # Over threshold + with pytest.raises(ValueError): + prepare_google_content_part(b"abc" * 10 * 1024 * 1024, "image/png") + +@pytest.mark.asyncio +async def test_upload_to_google_file_api(): + class _Part: + @staticmethod + def from_uri(file_uri, mime_type): + return {"uri": file_uri, "mime": mime_type} + + mock_client = MagicMock() + mock_upload_res = MagicMock() + mock_upload_res.uri = "gs://test-bucket/file-1" + mock_upload_res.mime_type = "image/png" + mock_client.files.upload.return_value = mock_upload_res + + # Mock google.genai.types.Part + with patch("google.genai.types.Part", _Part): + # 1. Successful upload without cache + res = await upload_to_google_file_api(b"data", "image/png", client=mock_client) + assert res == {"uri": "gs://test-bucket/file-1", "mime": "image/png"} + + # 2. Caching logic (hit and miss) + cache = ProviderMediaCache() + res_uncached = await upload_to_google_file_api(b"data", "image/png", cache=cache, client=mock_client) + assert res_uncached == {"uri": "gs://test-bucket/file-1", "mime": "image/png"} + + # Call again -> should hit cache and not call upload + mock_client.files.upload.reset_mock() + res_cached = await upload_to_google_file_api(b"data", "image/png", cache=cache, client=mock_client) + assert res_cached == {"uri": "gs://test-bucket/file-1", "mime": "image/png"} + mock_client.files.upload.assert_not_called() + + # 3. Default client creation test (mocking client = None) + mock_genai_client = MagicMock() + mock_genai_client.files.upload.return_value = mock_upload_res + with patch("google.genai.Client", return_value=mock_genai_client): + res_default_client = await upload_to_google_file_api(b"data_other", "image/png", client=None) + assert res_default_client == {"uri": "gs://test-bucket/file-1", "mime": "image/png"} + +def test_openai_helpers(): + search_tool = create_openai_file_search_tool(["file-1"]) + assert search_tool == { + "type": "file_search", + "file_search": { + "vector_store_ids": [] + } + } + + attachment = create_openai_file_attachment("file-2", ["file_search"]) + assert attachment == { + "file_id": "file-2", + "tools": [{"type": "file_search"}] + } diff --git a/tests/storage/test_init.py b/tests/storage/test_init.py new file mode 100644 index 0000000..abf3135 --- /dev/null +++ b/tests/storage/test_init.py @@ -0,0 +1,11 @@ +import pytest +import agentflow.storage + + +def test_storage_lazy_exports(): + assert agentflow.storage.make_agent_memory_tool is not None + assert agentflow.storage.make_user_memory_tool is not None + assert agentflow.storage.memory_tool is not None + + with pytest.raises(AttributeError): + _ = agentflow.storage.invalid_attribute_name_xxx diff --git a/tests/testing/test_quick_test.py b/tests/testing/test_quick_test.py index 7b8e0de..1eefb75 100644 --- a/tests/testing/test_quick_test.py +++ b/tests/testing/test_quick_test.py @@ -237,3 +237,58 @@ def test_extract_response_last_assistant(self): ] result = QuickTest._extract_response({"messages": msgs}) assert result == "final response" + + @pytest.mark.asyncio + async def test_with_tools_actual_execution(self): + """Test QuickTest.with_tools actual execution route.""" + # This will call QuickTest.with_tools, which routes through ToolNode + result = await QuickTest.with_tools( + query="Tell me the weather in SF", + response="The weather in SF is sunny", + tools=["get_weather"], + tool_responses={"get_weather": "sunny"} + ) + assert result is not None + assert result.final_response == "The weather in SF is sunny" + assert len(result.tool_calls) > 0 + assert result.tool_calls[0]["name"] == "get_weather" + + @pytest.mark.asyncio + async def test_multi_turn_fallback_no_state_context(self): + """Test multi_turn fallback when state context is not present.""" + # We can mock compiled.ainvoke to return a dict without state context + # to hit line 150 of quick_test.py + from agentflow.core.graph.compiled_graph import CompiledGraph + from agentflow.core.state import Message + from unittest.mock import AsyncMock, MagicMock, patch + + mock_compiled = MagicMock(spec=CompiledGraph) + mock_compiled.ainvoke = AsyncMock(return_value={ + "messages": [Message.text_message("User message", role="user"), Message.text_message("Response", role="assistant")] + }) + + with patch("agentflow.core.graph.StateGraph.compile", return_value=mock_compiled): + result = await QuickTest.multi_turn( + conversation=[("User message", "Response")] + ) + assert result.final_response == "Response" + assert len(result.messages) == 2 + + def test_extract_response_no_text_method_fallback(self): + """Test extract response fallback when message text attribute is not a method/callable.""" + # Create a mock message where hasattr(msg, "text") is False but hasattr(msg, "content") is True + msg = type("MockMsg", (), { + "role": "assistant", + "content": "Fall back to content string" + }) + result = QuickTest._extract_response({"messages": [msg]}) + assert result == "Fall back to content string" + + # Create a mock message where msg has no content but has role + msg2 = type("MockMsg", (), { + "role": "assistant" + }) + result = QuickTest._extract_response({"messages": [msg2]}) + # str(msg2) is returned + assert "MockMsg" in result + diff --git a/tests/utils/test_call_llm.py b/tests/utils/test_call_llm.py index 319ada3..90ecadc 100644 --- a/tests/utils/test_call_llm.py +++ b/tests/utils/test_call_llm.py @@ -9,6 +9,9 @@ from agentflow.core.llm.caller import ( _extract_responses_text, call_llm, + _call_google, + _call_openai_responses, + _call_openai_chat, ) @@ -157,3 +160,136 @@ def test_extract_falls_back_to_output_items(): def test_extract_returns_empty_when_no_text(): r = _make_response(output_text=None, output=[]) assert _extract_responses_text(r) == "" + + +# --------------------------------------------------------------------------- +# Direct private call implementations +# --------------------------------------------------------------------------- + +@pytest.mark.anyio +async def test_call_google_implementation(): + mock_client = MagicMock() + mock_response = MagicMock() + mock_response.text = " hello google " + mock_response.usage_metadata.prompt_token_count = 10 + mock_response.usage_metadata.candidates_token_count = 5 + mock_response.usage_metadata.cached_content_token_count = 3 + + mock_client.aio.models.generate_content = AsyncMock(return_value=mock_response) + + # 1. With system prompt, temperature, json mode + res = await _call_google( + mock_client, + "gemini-2.0-flash", + "user prompt", + system_prompt="sys prompt", + max_tokens=100, + temperature=0.5, + json_mode=True + ) + + assert res == ("hello google", 10, 5, 3) + mock_client.aio.models.generate_content.assert_called_once() + _, kwargs = mock_client.aio.models.generate_content.call_args + assert kwargs["model"] == "gemini-2.0-flash" + assert kwargs["contents"] == "user prompt" + config = kwargs["config"] + assert config.max_output_tokens == 100 + assert config.temperature == 0.5 + assert config.response_mime_type == "application/json" + assert config.system_instruction == "sys prompt" + + # 2. With cached content (system prompt should be ignored) + mock_client.aio.models.generate_content.reset_mock() + res = await _call_google( + mock_client, + "gemini-2.0-flash", + "user prompt", + system_prompt="sys prompt", + max_tokens=100, + temperature=0.5, + json_mode=False, + cached_content="cachedContents/abc123" + ) + + _, kwargs = mock_client.aio.models.generate_content.call_args + config = kwargs["config"] + assert config.cached_content == "cachedContents/abc123" + assert getattr(config, "system_instruction", None) is None + + +@pytest.mark.anyio +async def test_call_openai_responses_implementation(): + mock_client = MagicMock() + mock_response = MagicMock() + mock_response.output_text = " hello responses " + + mock_response.usage.input_tokens = 20 + mock_response.usage.output_tokens = 15 + mock_response.usage.input_tokens_details.cached_tokens = 7 + + mock_client.responses.create = AsyncMock(return_value=mock_response) + + res = await _call_openai_responses( + mock_client, + "gpt-4o-mini", + "user prompt", + system_prompt="sys prompt", + max_tokens=200, + temperature=0.7, + json_mode=True, + extra_param="extra_val" + ) + + assert res == ("hello responses", 20, 15, 7) + + mock_client.responses.create.assert_called_once() + _, kwargs = mock_client.responses.create.call_args + assert kwargs["model"] == "gpt-4o-mini" + assert kwargs["input"] == "user prompt" + assert kwargs["max_output_tokens"] == 200 + assert kwargs["temperature"] == 0.7 + assert kwargs["instructions"] == "sys prompt" + assert kwargs["text"] == {"format": {"type": "json_object"}} + assert kwargs["extra_param"] == "extra_val" + + +@pytest.mark.anyio +async def test_call_openai_chat_implementation(): + mock_client = MagicMock() + mock_response = MagicMock() + + choice = MagicMock() + choice.message.content = " hello chat " + mock_response.choices = [choice] + + mock_response.usage.prompt_tokens = 30 + mock_response.usage.completion_tokens = 25 + mock_response.usage.prompt_tokens_details.cached_tokens = 12 + + mock_client.chat.completions.create = AsyncMock(return_value=mock_response) + + res = await _call_openai_chat( + mock_client, + "gpt-4o-mini", + "user prompt", + system_prompt="sys prompt", + max_tokens=300, + temperature=0.8, + json_mode=True, + another_param="another_val" + ) + + assert res == ("hello chat", 30, 25, 12) + + mock_client.chat.completions.create.assert_called_once() + _, kwargs = mock_client.chat.completions.create.call_args + assert kwargs["model"] == "gpt-4o-mini" + assert kwargs["messages"] == [ + {"role": "system", "content": "sys prompt"}, + {"role": "user", "content": "user prompt"} + ] + assert kwargs["max_tokens"] == 300 + assert kwargs["temperature"] == 0.8 + assert kwargs["response_format"] == {"type": "json_object"} + assert kwargs["another_param"] == "another_val"