Framework and Language Coverage

agent_architecture detects common agent orchestration patterns and applies configurable language filtering.

Framework Detection Coverage

The detector family covers these framework patterns:

Framework	Languages	Typical Anchors
LangGraph	Python, TypeScript	`StateGraph(`, `CompiledGraph`, `.compile()`, `invoke(`
CrewAI	Python, TypeScript	`Agent(`, `Task(`, `Crew(`, `.kickoff(`
LangChain Agents	Python, TypeScript	`AgentExecutor`, `create_react_agent`, `createReactAgent`, `.run(`, `.call(`, `.invoke(`
AutoGen	Python	`AssistantAgent(`, `GroupChat(`, `GroupChatManager(`, `initiate_chat(`
OpenAI Agents SDK	Python, TypeScript	`Runner.run(`, `Runner.run_sync(`, `Runner.run_streamed(`, `@openai/agents`
Pydantic AI	Python	`pydantic_ai.Agent(`, `.run_sync(`, `.iter(`
DSPy	Python	`dspy.ChainOfThought`, `dspy.ReAct`, `dspy.Module`
Semantic Kernel	Python	`semantic_kernel`, `Kernel(`, `kernel.invoke(`
LlamaIndex Agents	Python	`ReActAgent(`, `AgentRunner(`
Amazon Bedrock Agents	Python, TypeScript	`bedrock-agent-runtime`, `invoke_agent(`, `InvokeAgentCommand`
Google ADK	Python	`LlmAgent(`, `google.adk.agents`
Haystack Agents	Python	`haystack.components.agents`, `PipelineTool`, `agent.run(`
Vercel AI SDK	TypeScript	`generateText(`, `streamText(`, `useChat(`, `tool(`
Mastra	TypeScript	`mastra/core`, `createTool(`, `Agent.stream(`

Call sites are deduplicated by (file, line) to avoid overcounting. Generic invoke/compile patterns are counted only when a framework anchor exists in the same file.

Language Filtering Behavior

The optional metrics[].config.languages filter is extension-based.

If languages is empty, no language filter is applied.
If set, only call sites in matching extensions are kept.

Supported language values and mapped extensions:

`languages` value	Extensions
`python`	`.py`
`typescript`, `ts`, `javascript`, `js`	`.ts`, `.tsx`, `.js`, `.jsx`, `.mjs`, `.cjs`

Example:

metrics:
  - id: agent_architecture
    config:
      languages: ["python", "typescript"]

AST vs Pattern Fallback

At a high level:

Python and TypeScript call-site detection uses AST parsing when parse succeeds.
If parsing fails, detection falls back to pattern-based matching.

For loop-guard checks, Python can use same-function AST context to improve guard detection confidence.

Known Limitations and FP/FN Caveats

Heuristic pattern matching may produce false positives in unrelated helper code with similar naming.
Framework wrappers and internal abstraction layers can hide call sites and cause false negatives.
Extension-based filtering depends on file names; non-standard extensions can be excluded unintentionally.
Schema validation detection focuses on common conventions (Pydantic/Zod/JSON-like patterns) and may miss custom validators.
Confidence can vary by detector mode (AST-backed vs line/pattern heuristics).

Practical Recommendations

Start without a languages filter to establish baseline visibility.
Add languages only if you need to scope CI runtime or ownership boundaries.
Pair policy thresholds with finding review before strict enforcement.