SDK Reliability Audit

AI guesses.We check.

voice_agent.py · written by an AI assistantretell-sdk 6.2.19
1number = client.phone_number.create(
2    inbound_agent_id="agent_123",
3    area_code=415,
4)
TypeError: create() got an unexpected keyword argument 'inbound_agent_id'
Checked against the installed package, never run.
fix: inbound_agents=["agent_123"]

Your developers ask an AI assistant before they read your docs. We give it 30 real tasks for your product and prove every call it gets wrong, against your exact Python SDK.

See a real failure

One wrong word

Asked to route calls to an agent, the assistant wrote a keyword the SDK has never had. Python stops with a TypeError; your developer opens a ticket or tries a competitor.

What the assistant wroteinbound_agent_id
What the SDK takesinbound_agents

From our pilot on a voice AI company's Python SDK, version 6.2.19 (vendor anonymised).

The same mistake appeared in a second method, so one sentence in the docs fixes both.

How the audit works

First, a real task for your product.

Create a phone number with area code 415 and route its inbound calls to my agent.
The assistant's answerfrom memory, no docs
number = client.phone_number.create(
    inbound_agent_id="agent_123",
    area_code=415,
)

Then, proof against your exact version.

Clean install of the release you choose; every import, attribute and keyword checked. Nothing is run, no keys needed.
Verdictretell-sdk 6.2.19
phone_number.create() has no keyword inbound_agent_id
nearest: inbound_agents, outbound_agents
result:  TypeError

Last, the report and the fix.

Every failure with its line, the call that exists instead, and the sentence to add to your docs and llms.txt.

Audit summary: voice AI SDK 6.2.19

Claude Sonnet 5.5, from memory · 30 tasks · checked against a clean install

VerdictCallsMeaning
TypeError2Keyword argument that does not exist
Deprecated0Marked deprecated in the package
Unverified18Could not be proven either way, never counted as a failure
Verified94Proven to exist with this signature
Fix for llms.txt: phone_number.create and phone_number.update route calls with inbound_agents and outbound_agents (lists), not inbound_agent_id or outbound_agent_id.

No false alarms

false failures across 5,652 lines of correct code in our precision test, November 2026. A failure is reported only when Python itself would refuse the call.

Checker tests, November 2026

TestPlantedFoundFalse
stripe 16.1.0, 5 snippets220
openai 3.46.0, 6 snippets110
Precision sweep, 5,652 lines of correct code000
Deprecation markers flagged1431430

One SDK, one price

Pay by card, reply with your package name, and the report arrives in three business days. If it is not useful to you, we refund it.

Delivered in 3 business daysReport · fix list · tasks, answers and findings as JSON
$490once
  • 30 real tasks for one Python SDK
  • Every failure proven against your exact version
  • Fix sentences for your docs and llms.txt
  • One re-run within 60 days
Order the audit

Questions

Which assistant do you test with?

Claude Sonnet 5.5, answering from memory without your docs, the way most developers first meet your SDK. The report names the assistant and model version.

Do you call our API or need keys?

No. The code is parsed, never run. We install your published package and inspect it, so no keys, no traffic and none of your data.

What if nothing fails?

You get the verified calls, the unverified list and all 30 tasks: proof you can quote in your docs. If that is not worth it to you, you get a refund.

What do you need from us?

The PyPI package name and version, and any workflows you want covered. Nothing else.

Do you cover TypeScript or Go?

Python only for now. Tell us which SDK you need next; TypeScript is the next one we build.