Give your coding agent a compliance sub-agent

The pattern that always works: a sub-agent step

One honest constraint first: the endpoint is text-only and rejects tool and function definitions. Tool-using agent modes cannot run on it directly. So the universal shape is the one below: your agent keeps its tools and its general model, and calls the compliance step with plain messages, pinning the frameworks for the task.

from openai import OpenAI

client = OpenAI(
    base_url="https://api.ismscopilot.com/v1",
    api_key="sk-isms-...",
)

resp = client.chat.completions.create(
    model="isms-thinking",
    messages=[
        {"role": "user", "content": "Does this PR description affect ISO 27001 Annex A access controls? Name the controls."}
    ],
    extra_body={"ismscopilot": {"frameworks": ["ISO_27001"]}},
)
print(resp.choices[0].message.content)

The three facts every OpenAI-compatible client needs: base URL https://api.ismscopilot.com/v1, key sk-isms from platform.ismscopilot.com/keys, model alias isms-fast or isms-thinking (plus -eu twins). Wrap that call in a script your harness can run, and the sub-agent exists.

Pin the framework, do not hope for detection

Auto-detection scans the user messages and the last assistant turn, not the system prompt. In a harness, your instructions live in the system prompt, so auto can miss. Pin instead: send the ismscopilot extension with exact catalog ids, up to eight per call.

curl https://api.ismscopilot.com/v1/chat/completions \
  -H "Authorization: Bearer sk-isms-..." \
  -H "Content-Type: application/json" \
  -d '{
    "model": "isms-fast",
    "messages": [
      {"role": "user", "content": "Which ISO 27001:2022 Annex A control covers acceptable use of information and assets? One line."}
    ],
    "ismscopilot": {"frameworks": ["ISO_27001"]}
  }'

Verified on 2026-08-25, that request returns the right control (A.5.10) with the disclosure on the response: header x-isms-frameworks: ISO_27001 and an ismscopilot object listing the injected modules and their knowledge size. Log the disclosure per request as evidence of which modules grounded each step. Valid ids come from GET /v1/frameworks.

Honest limits

1. 1.Text in, text out

   The endpoint speaks the OpenAI chat completions subset: text messages, streaming, usage. It does not accept tool or function definitions, JSON mode, logprobs, n above 1, or multimodal parts. The sub-agent gets grounded generation, not tool use. Your harness keeps its own tools.

2. 2.Detection is name-level, and it does not read your system prompt

   Auto mode scans the user messages and the last assistant turn. A bare control id like 5.23 injects nothing. Harness instructions usually live in the system prompt, which auto does not scan at all. In an agent loop, pin the frameworks explicitly.

3. 3.Thinking has a floor

   Thinking aliases are for harder questions and floor max_tokens at 1024 when a lower value is sent, surfaced in response headers. Fast answers cost the same unit rate and are usually enough for a review step.

4. 4.Guidance, not an audit opinion

   Answers are educational compliance guidance with the module disclosed. Risk acceptance and anything a regulation assigns to a person or a management body stay with you.

Frequently asked questions