Skip to content

Quick start

60-second flow: install → first detection

pip install 'valanistack[anthropic]'
export ANTHROPIC_API_KEY="sk-ant-..."

from vstack.aar import AnthropicClient, AARAnalyzer, AgentTrace, TraceStep

trace = AgentTrace(
    goal="Refactor the auth module to use JWTs.",
    steps=[
        TraceStep(type="tool_call",   content="edit_file(path='auth/middleware.py')"),
        TraceStep(type="observation", content="session-middleware test failures"),
        TraceStep(type="output",      content="Created JWT tokens but broke sessions."),
    ],
    outcome="Auth module half-migrated; session middleware broken.",
    success=False,
)

aar = AARAnalyzer(AnthropicClient(), mode="standard").run(trace)
print(aar.to_markdown())

That's the foundational diagnostic: an After-Action Review. The output is a structured detection with lessons[], next_steps[], trace_quality_audit, plus a list of recommended downstream patterns to chain into.

From there

Want the full failure-attribution chain? AAR → Lewin → downstream is the vstack-post-incident skill (or just call the patterns directly):

from vstack.lewin import LewinAttributionDetector, AgentFailureTrace, FailureStep

lewin = LewinAttributionDetector(AnthropicClient(), mode="standard").run(
    AgentFailureTrace(
        agent_id="qa-bot",
        model_name="claude-opus-4-7",
        task=trace.goal,
        steps=[FailureStep(type=s.type, content=s.content) for s in trace.steps],
        outcome=trace.outcome,
        success=False,
        initial_attribution="model bug",
    )
)
print(lewin.dominant_locus)   # 'internal' / 'environmental' / 'interactional'

Want vstack inside your AI client? Set up the MCP server:

pip install 'valanistack[anthropic,mcp]'
vstack-mcp config-snippet claude-desktop   # prints a paste-ready JSON block

Paste the block into ~/Library/Application Support/Claude/claude_desktop_config.json (macOS), restart Claude Desktop, ask: "Use the AAR pattern on this trace…"

Want vstack inside a LangChain agent?

pip install 'valanistack[anthropic,langchain]'

from vstack.adapters.langchain import as_langchain_tools
tools = as_langchain_tools()  # 34 StructuredTool instances
# hand them to your LangChain agent

See Framework adapters for the full set.

Next steps