· v6.2.0 shipped · nothing crosses unseen ·

styxx.profile
py-spy for LLM reasoning

langsmith shows you that your agent failed.
the cognitive profiler shows you why. one decorator. every failure mode localized to the step it happened.

· 01 · install + use ·

one decorator. that's it.

Wrap your agent function. Get back the result and a cognitive profile. No config, no auth, no API to join.

# pip install -U styxx
import styxx

@styxx.profile
def my_agent(task):
    return run_langchain_agent(task)

result, p = my_agent("summarize this contract")

print(p.summary)       # one-line verdict
print(p.faults)        # step-localized faults
p.to_html("run.html")   # flamegraph for humans
p.to_json("run.json")   # langsmith / datadog export

Works on openai, anthropic, huggingface, langchain, llamaindex, crewai, autogen, langsmith, langfuse, and raw logprobs. Fails open — if styxx can't read vitals, your agent returns its normal response unchanged.

· 02 · seven fault kinds ·

every cognitive failure, localized to the step.

Other tools give you traces. Profiler gives you causes. Seven canonical fault kinds, each with a severity score and the step index that produced it.

drift
wrong tool / wrong args
categories: arg_swap · tool_arg_drift · tool_confab. flagged when confidence > 0.5.
confabulation
model made it up
categories: confab · hallucination · fabrication. the moment your agent confidently lied.
refusal
model refused
strong-refusal mode only (conf > 0.8). cascaded refusals across steps show up as faults.
sycophant
yes-man detector
your crew has 4 agents and 1 is agreeing with everything. find the yes-man.
phase transition
mode flip mid-run
adjacent steps shift category — the moment cognition pivots. novel in any production profiler.
low trust
the readout itself is shaky
trust < 0.30. the instrument is telling you the instrument isn't sure either. honest about its own uncertainty.
incoherence
c-axis collapse
cross-phase coherence < 0.30. the model's commitment decayed between reasoning and output.
backed by science. Hallucination AUC 0.998 on HaluEval-QA. Refusal AUC 0.976 on GPT-4 out-of-family. Tool-drift AUC 0.943 on BFCL v3. All cross-validated across 8 benchmarks with published failure modes. Zenodo deposit · papers/
· 03 · drops into your stack ·

three export formats. no lock-in.

format method what it gives you
HTML flamegraph p.to_html(path) self-contained, zero external assets. dark terminal aesthetic. screenshot-ready for your next postmortem.
LangSmith trace p.to_langsmith() parent chain + child llm spans, styxx vitals as flat metadata. drop into client.create_run(...).
Datadog spans p.to_datadog() apm-style spans, category + gate as tags, confidence + trust as metrics. auto-faulted spans tagged.
raw dict p.to_dict() everything, serializable. pipe it wherever.
· 04 · $STYXX ·

there's a token. here's the honest version.

$STYXX is a Solana Token-2022 mint tied to this project. Live on-chain since the early days of fathom lab. Not a sale, not an ICO. We're not airdropping — supply is fixed on pump.fun, we don't hold distributable allocations.

contract address · Solana Token-2022
Dxw3u4KxN32KpSdHSq4TkwjfMPJTPeosa22JXN15pumpclick to copy
program
Token-2022
decimals
6

two utilities. both enforceable.

utility what it does when
atlas pro gate hold 100k $styxx in your connected wallet → unlock closed-model profiling (via the anthropic_hack pipeline), layer-level fault localization, and weekly-updated calibrations for 500+ models. no lock, just balance check. 2–3 weeks
revenue → buy-back 5% of styxx commercial licensing revenue used to buy $styxx on the open market and burn. every burn tx hash logged in CHANGELOG.md, publicly verifiable. starting with first license

no airdrop. no presale. no unlocks. no team allocation. The only way in is to buy on-market. The only way to hold is to believe the product keeps shipping. We consider that a feature.

· 05 · atlas pro priority access ·

get verified first.

When atlas pro rolls out, early-access slots go first to developers who've actually run the profiler. Drop your email + wallet + a profile JSON (from p.to_json()). We verify the profile is real, you skip the line.

we check balance at rollout — doesn't have to be funded today.
validates locally — must include at least one step and parse as valid styxx profile. this is your proof you ran the tool.