qa/acceptance/README.md

Rama main · réplica de lectura del repositorio original.

# Relmio release acceptance workflow

`relmio-release-acceptance.json` is an n8n 2.40.7 workflow export for the live
checks in the `release-qa` skill (`.agents/skills/release-qa/SKILL.md`). It
calls the ChatGPT sidecar through each n8n node Relmio supports, plus the image
add-on and the local model. A release run executes it once by hand and reads
each node's output.

It is the standard QA workflow for every self-hosted n8n that uses Relmio: the
Hostinger VPS, a local Docker n8n, or any other setup. Each n8n keeps one copy,
named "Relmio release acceptance", and every QA run reuses it. Never import a
second copy into an n8n that already has one.

The export is inactive and contains no credentials. Nodes refer to
credentials by name only, so nothing secret is in this file.

## Nodes

| Node | Calls | Expected output |
| --- | --- | --- |
| HTTP Request: Responses | `POST /v1/responses` | `http responses ok` |
| HTTP Request: Chat Completions | `POST /v1/chat/completions` | `http chat ok` |
| AI Agent, with Agent Chat Model (Responses API on) and Calculator | Responses route with one tool call | `391`, and `intermediateSteps` shows a Calculator action whose observation is `391` (an answer without that step fails) |
| Basic LLM Chain, with Chain Chat Model (Responses API on) | Responses route | `chain ok` |
| OpenAI: Message a Model (node 2.3) | Responses route | `openai node ok` |
| OpenAI: Generate an Image (node 2.3, `gpt-image-2`) | `POST /v1/images/generations` | one binary image; image add-on only (VPS or local) |
| Local Model Chain, with Local Chat Model (Responses API off) | `http://n8n-local-model:11434/v1` | `4`; only where the local model is installed |

`Settings` holds the text model ID, the local model ID, the sidecar Base URL
(`sidecarBaseUrl`) and the workflow version (`acceptanceVersion`). The
defaults are `gpt-6-astra`, which is also the model the n8n AI Assistant uses,
`qwen3:0.6b` and `http://n8n-openai-oauth:10531/v1`. Every node reads them
from `Settings`. Raise `acceptanceVersion` whenever this file changes.

Before a run, open the Relmio wizard's **Check installed account** for the
target sidecar and set `model` to a model it shows as Ready. On 2026-10-07 the
owner's catalog also had `gpt-6.1-sol`, `gpt-6-sol` and `gpt-6-luna`. Set
`localModel` to the model the local model status shows as ready.

Each check node is set to continue on error, so one failure does not hide the
others. That also means a failed node can look finished. A check passes only
when its output shows the expected text. An `error` field is a failure. Image
and local model failures are expected where those add-ons are not installed;
record them as skipped with the reason.

## Credentials

Both credentials are n8n **OpenAI** credentials (`openAiApi`).

| Name | Base URL | API key | Used by |
| --- | --- | --- | --- |
| `OpenAI Relmio` | `http://n8n-openai-oauth:10531/v1` | The one-time Relmio bearer from the wizard | Every node except Local Chat Model |
| `Relmio Local Model` | `http://n8n-local-model:11434/v1` | `local-only` (ignored) | Local Chat Model |

The HTTP Request nodes use the predefined `OpenAI` credential type, so the
bearer stays in the credential and never appears in the URL or body. Their URLs
come from `sidecarBaseUrl`. The default is the sidecar's Docker hostname, which
is the same on the Hostinger VPS and a local Docker n8n. On another self-hosted
n8n, set `sidecarBaseUrl` and the `OpenAI Relmio` Base URL to the address that
n8n reaches the sidecar at, and the `Relmio Local Model` Base URL to the local
model's address.

## Find, reuse or update

1. In n8n, open **Workflows** and search for "Relmio release acceptance".
2. If it exists, reuse it. Open `Settings` and compare `acceptanceVersion`
   with this file. When they match, go straight to the run. When the file is
   newer, or the workflow predates `acceptanceVersion`, update it in place:
   open the workflow, select all nodes, delete them, then choose **Import >
   From URL** (or **From File**) in the same workflow's menu with this file.
   The workflow keeps its ID, name and execution history.
3. Only when no copy exists, create a workflow and import this file into it
   the same way. Keep it inactive. It has only a manual trigger, so activating
   it would do nothing useful.
4. The export carries credential names but no IDs. n8n links an existing
   credential with the same name, but check every node with a credential
   warning and pick the credential from the table above: both HTTP Request
   nodes, both OpenAI nodes, Agent Chat Model and Chain Chat Model use `OpenAI
   Relmio`, and Local Chat Model uses `Relmio Local Model`. If that n8n names
   them differently, pick the matching credential by its Base URL; do not
   rename the owner's credentials.
5. Set the model IDs, and `sidecarBaseUrl` where it differs, in `Settings`,
   then save.

Creating or updating the workflow in the owner's n8n is a separate step that
needs the owner's confirmation. It changes nothing else in n8n and does not
restart it.

## Plan usage

One run sends six text requests to the ChatGPT plan (the AI Agent makes two
because of the tool call) and one image request against the Codex image
limits when the add-on is on. The local model uses no plan. State these
numbers in the release evidence.