Content capture and privacy

By default the SDK sends only cost metadata: tokens or billing units, model, latency, and tags. Prompt and completion text are not captured.

Hosted OTLP is stricter in this release: it always discards prompt/completion, system instruction, tool argument/result, and binary content before request rows are published. The SDK content-capture setting does not enable hosted OTLP content capture. If OpenRouter Privacy Mode is off, content still traverses the receiver before being discarded; enable it when content must never be transmitted.

Opt-in capture

Organizations on Pro or Enterprise can enable sampled content capture from Settings. Once enabled:

• A sampled subset of prompts and completions is captured automatically.
• Samples are text-only (images, audio, and binary parts are stripped).
• Secrets and common personal identifiers are redacted in your process before a sample is sent (see What is redacted), and content is encrypted at rest.
• Org owners can delete captured content from Settings.

There is nothing to call in application code to turn capture on. The Settings opt-in is always required.

What is redacted

Before a sample leaves your process, the SDK replaces the items below with a typed placeholder. The Python and JavaScript SDKs apply the same rules to the prompt, the outcome, and every string inside tool calls and tool results.

Names, quotes and auth schemes around a secret stay readable, for example Authorization: Bearer [REDACTED_SECRET] or DB_PASSWORD=[REDACTED_SECRET].

SuperPenguin applies the [REDACTED_SECRET] rules again when a sample arrives, before it is encrypted, so samples from older SDK versions get the same secret redaction. Personal data is redacted only in your process.

These are not redacted:

• Names, street addresses, dates of birth and other free-text personal data. Detecting them reliably needs a language model or entity recognizer, which the SDK does not bundle.
• National ID numbers other than US SSNs, nine-digit numbers without dashes or spaces, and phone numbers without a + outside the North American formats above.
• Secrets without a recognizable prefix or assignment, for example a random token pasted on its own, an Azure OpenAI key (32 hex characters) without a name such as api-key: in front of it, or the secret half of a Google Cloud Storage HMAC key.
• Metadata values such as customer_id, session_id, feature and prompt_key. They are sent as you set them, so do not put email addresses or other personal data in them.

A version string shaped like a public IPv4 address (release 1.2.3.4) is masked as an IP address.

Add your own rules

Pass a redactor function to add rules of your own. It runs after the built-in rules, on the prompt text, on the outcome text, and on the tool calls and tool results serialized as one JSON string. A sample it changes is marked as redacted.

import re
import superpenguin as sp

sp.configure_content_capture(
    redactor=lambda text: re.sub(r"\bACME-\d{6}\b", "[ACCOUNT]", text),
)

import { configureContentCapture } from "@superpenguin/js";

configureContentCapture({
  redactor: (text) => text.replace(/\bACME-\d{6}\b/g, "[ACCOUNT]"),
});

Each field is stored up to 64,000 bytes. Longer text is cut after redaction, and the sample records which fields were cut.

Group multi-turn conversations

Set two metadata keys:

• session_id: one conversation
• turn_id: one turn (reuse across a tool round-trip)

Each model call is its own row. A turn that runs model → tool → model spans two rows; reuse turn_id to group them.

with sp.metadata({
    "session_id": conversation_id,
    "turn_id": turn_id,
    "customer_id": "cust_acme_123",
}):
    first = client.chat.completions.create(...)
    # run tool locally, then:
    final = client.chat.completions.create(...)

Narrow capture from the SDK

SDK helpers can only reduce capture, never enable it:

sp.configure_content_capture(
    allow=True,            # False vetoes all capture from this process
    record_prompt=True,    # False drops inputs
    record_outcome=True,   # False drops outputs
    sample_rate_multiplier=1.0,
)

import { configureContentCapture } from "@superpenguin/js";

configureContentCapture({
  allow: true,
  recordPrompt: true,
  recordOutcome: true,
  sampleRateMultiplier: 1,
});