@STYXX.PROFILE · v6.2.0+

py-spy for LLM reasoning.

One decorator on your agent function. Per-step cognometric readout — drift, confabulation, refusal, sycophancy, phase transition, low trust, incoherence — localized to the step that produced them. langsmith tells you the trace broke. styxx.profile tells you why.

pip install -U styxx docs

01 · usage

One decorator. Any LLM-using function works inside.

import styxx
from styxx import OpenAI   # raw openai, langchain, crewai, autogen — all work

@styxx.profile
def my_agent(task):
    client = OpenAI()
    r = client.chat.completions.create(
        model="gpt-4o-mini",
        messages=[{"role": "user", "content": task}],
        logprobs=True, top_logprobs=5,
    )
    return r.choices[0].message.content

result, profile = my_agent("summarize this report")

for step in profile.steps:
    print(step.cognometric_fingerprint)
    # {sycophancy: 0.04, drift: 0.18, hallucination: 0.02, ...}

02 · what it gives you

Per-step localization. Every fault, attached to its origin.

DRIFT

Goal drift across turns

The agent moved away from its anchor. Localized to the turn where divergence began.

CONFAB

Confabulation

Fluent claim with no grounding signal. Per-token attribution of which claim broke.

REFUSE

Refusal

The model declined. Did it decline because it should — or because it confused capability with policy?

SYCOPH

Sycophancy

Yielding without evidence. K=1 superlative-density signature, attached to the step.

PHASE

Phase transition

The model crossed a discrete state boundary mid-trace. Refusal → answer, retrieval → reasoning, etc.

TRUST

Low trust

Cognometric reward score per response. Drop-in for cogn-RLHF training reflection.

One line. Cognitive vitals on every agent.

pip install -U styxx — then add @styxx.profile above any function. Works with raw openai, langchain, crewai, autogen, custom code.

install on pypi github