aiyara 0.2.0

Agent failure monitoring SDK — trace capture with client-side PII scrubbing

Aiyara — Agent Failure Monitoring SDK

Aiyara is agent failure monitoring with system-prompt-driven constraint extraction. Paste your system prompt at app.aiyara-universe.com, we extract structural constraints (tool ordering, argument rules, cross-turn dependencies) as a deterministic policy, and every trace is checked against that policy plus pre-built harness detectors (doom loops, retry amplification, context overflow, cascades). Violations surface as issues with evidence, not LLM-judge guesses.

This Python SDK captures traces from your agent in production and ships them to the Aiyara ingest service for evaluation. Capture is non-blocking, batched, and PII is scrubbed client-side before any bytes leave your process.

Quick start

pip install aiyara

import aiyara

aiyara.init(api_key="ak_...", auto_instrument=True)

That's it. With auto_instrument=True, the SDK patches the supported client libraries below and every agent run is captured and sent to Aiyara automatically.

Get an API key

1. Visit app.aiyara-universe.com.
2. Sign up (Supabase auth — no card required for the free tier).
3. Paste your system prompt. The wizard extracts your policy rules and lets you review them.
4. Issue an ak_* API key for your project. Use it in aiyara.init(api_key=...).

Auto-instrumentation

Supported frameworks (patched on aiyara.init(auto_instrument=True)):

• openai — Chat Completions, Responses, tool calls
• langgraph — graph nodes, tool calls, state transitions
• openai_agents — Agents SDK runs

You can also drive capture manually:

# Single capture
aiyara.capture(
    messages=[{"role": "user", "content": "Hello"}],
    tool_calls=[{"name": "search", "arguments": {"q": "test"}, "result": "ok"}],
    model="gpt-4",
)

# Context-managed trace
with aiyara.trace(session_id="user-123") as t:
    t.record_message(role="user", content="Find flights")
    t.record_tool_call(name="search_flights", arguments={}, result={})

aiyara.shutdown()  # flush pending traces

PII scrubbing

PII scrubbing happens client-side, before transport. Raw user content never leaves your process unless explicitly opted in. The default scrubber redacts emails, phone numbers, and other common PII patterns; pass pii_field_paths=[...] to extend it.

aiyara.init(
    api_key="ak_...",
    scrub_pii=True,                          # on by default
    pii_field_paths=["messages[*].content"], # extend default scrubbing
)

What you get

• Per-constraint verdicts. Aiyara tells you which rule fired, on which step, with the trace evidence — not a number from a judge.
• Cascade detection. Root-cause violations are linked to their downstream failures via depends_on graphs.
• Pre-built harness detectors. Doom loops, context overflow, retry amplification, error cascades — zero configuration required.
• Deterministic monitoring. Pure-Python checks. Compile policy once, evaluate every trace, zero LLM calls on the hot path.

License

MIT — see LICENSE for details.

Links

• Homepage: www.aiyara-universe.com
• Dashboard: app.aiyara-universe.com
• Source: github.com/Aiyara-Universe/Testbench
• Changelog: sdk/CHANGELOG.md
• Issues: github.com/Aiyara-Universe/Testbench/issues