> ## Documentation Index
> Fetch the complete documentation index at: https://docs.epode.ai/llms.txt
> Use this file to discover all available pages before exploring further.

# How Epode works

> The company-side integration, customer-agent handoff, and feedback path.

Epode has two participants with different responsibilities:

| Participant           | Responsibility                                                                          |
| --------------------- | --------------------------------------------------------------------------------------- |
| Your company          | Adds the server-side SDK to selected product responses and keeps the product key secret |
| Your customer's agent | Uses the product, interprets the bounded feedback contract, and may submit one report   |

The SDK signs a capability locally. It does **not** call Epode before returning your product response.
Opportunity telemetry is delivered later through a bounded background queue.

## What the capability can do

The short-lived capability can submit one report for one interaction. It contains only a protocol
version, interaction ID, issue/expiry timestamps, and nonce. It does not contain customer context,
product content, a prompt, or an identity.

## What a report contains

```json theme={null}
{
  "summary": "The result answered the question, but one field required verification.",
  "impact": "helped_with_friction",
  "confidence": 0.91,
  "findings": [
    { "kind": "strength", "topic": "relevance", "detail": "The top result was useful." },
    { "kind": "gap", "topic": "freshness", "severity": "major", "detail": "One field was stale." }
  ],
  "workaround": { "used": true, "detail": "The agent verified the field in current docs." }
}
```

This is intentionally more expressive than success/partial/failure. One experience can contain a
strength, friction, uncertainty, gap, and workaround at the same time.

## Failure isolation

If Epode is slow or unavailable:

* your product response is not delayed or failed;
* queued opportunity telemetry may be retried or dropped;
* a feedback submission receives a retryable failure;
* no prompt, transcript, credential, or raw product payload is sent as telemetry.