groq.md

Groq — .klickd v3.4 Integration

SDK: groq-python · pip install groq

Groq's API is OpenAI-compatible. Recommended models for .klickd payloads:

Model	Use case
llama-3.3-70b-versatile	General purpose, strong context adherence
qwen/qwen3-32b	Preferred for injection_resistance_level: strict

⚠️ Avoid llama-3.1-8b-instant when injection_resistance_level is moderate or strict (confirmed masked compliance vulnerability — SPEC §23.2).

import json
from groq import Groq

client = Groq()  # uses GROQ_API_KEY env var

with open("profile.klickd") as f:
    klickd = json.load(f)

system_prompt = klickd.get("user_preferences", "")

ctx = klickd.get("context", {})
if ctx.get("current_state"):
    system_prompt += f"\n\nCurrent state: {ctx['current_state']}"
if ctx.get("resume_trigger"):
    system_prompt += f"\n{ctx['resume_trigger']}"

response = client.chat.completions.create(
    model="llama-3.3-70b-versatile",
    max_tokens=2048,          # Set explicitly — payloads >1500 tokens risk truncation
    messages=[
        {"role": "system", "content": system_prompt},
        {"role": "user", "content": "Let's continue."}
    ]
)

print(response.choices[0].message.content)

Truncation: Groq models may silently truncate responses on payloads exceeding ~1500 tokens.
Always set an explicit max_tokens value when injecting .klickd payloads.
Strip all _-prefixed fields (e.g. _benchmark) before injection.