Instance AI E2E Tests

End-to-end tests for the Instance AI feature, using recorded LLM responses replayed through a MockServer proxy. Tests run without an API key in CI while producing real database state for full frontend verification.

Table of Contents

• Architecture Overview
• How Recording Works
• How Replay Works
• The ID Problem and Tool Wrapping
• Two-Tier Tool Strategy
• Trace Format
• Proxy Expectations
• Test Infrastructure
• Running Tests
• Re-Recording Tests
• Robustness and Maintenance
• Adding New Tests
• Troubleshooting

Architecture Overview

Instance AI tests exercise a multi-agent LLM system that builds and executes n8n workflows. Each test sends a chat message, the LLM orchestrates tool calls (build-workflow, run-workflow, etc.), and the test asserts on the resulting UI state.

The challenge: LLM API calls are expensive, non-deterministic, and unavailable in CI. The solution is a record/replay architecture with two layers:

                        Recording (local dev)              Replay (CI)
                        =====================              ===========

LLM calls           →  Real Anthropic API              →  MockServer returns
                        (proxied + recorded)               recorded responses

Tool execution      →  Real execution                  →  Real execution
                        (input/output → trace.jsonl)       with ID remapping

Frontend            →  Real DB state                   →  Real DB state
                        (preview iframe works)             (preview iframe works)

Why Not Pure Mocking?

The Instance AI frontend loads workflow previews via an iframe that fetches real workflow data from the n8n API. If tools are fully mocked, the database has no workflows, and the preview shows nothing. By executing tools for real during replay, the database contains actual workflows, executions, and credentials that the frontend can render.

How Recording Works

When running locally with a real ANTHROPIC_API_KEY:

1. Proxy captures LLM traffic: All HTTP from the n8n container routes through a MockServer proxy (HTTP_PROXY/HTTPS_PROXY). Anthropic API calls (POST /v1/messages) are intercepted and recorded.

2. Tool calls are traced: Every tool invocation is recorded to a TraceWriter with the tool name, agent role, input, and output. Suspend/resume events (human-in-the-loop approvals) are recorded separately.

3. Fixture teardown saves both artifacts:

   • Proxy expectations → individual JSON files per HTTP exchange
   • Tool trace → trace.jsonl (one JSON object per line)

Both are saved under expectations/instance-ai/<test-slug>/.

How Replay Works

In CI (no API key):

1. Fixture setup loads artifacts: The instanceAiProxySetup auto-fixture reads proxy expectations and trace events from disk, uploads them to MockServer and the n8n container respectively.

2. LLM calls hit MockServer: The proxy returns pre-recorded responses in sequence. Each expectation fires once (remainingTimes: 1), except the last /v1/messages expectation which is unlimited (fallback for any extra calls).

3. Tools execute for real with ID remapping: When the LLM response triggers a tool call, the tool runs against the real database. But IDs from the recorded session won't match the current run's IDs. The IdRemapper translates between them.

4. Frontend works normally: Since tools produced real DB state with real IDs, the workflow preview iframe loads actual workflows.

The ID Problem and Tool Wrapping

This is the core challenge that motivated the trace replay infrastructure.

The Problem

Consider a test that builds and runs a workflow:

Recording session:
  build-workflow → { workflowId: "5" }
  run-workflow({ workflowId: "5" }) → { executionId: "exec-100" }

Replay session:
  build-workflow → { workflowId: "12" }     ← different auto-increment ID
  run-workflow({ workflowId: "5" }) → ERROR  ← LLM still says "5" (from recorded response)

The LLM response is pre-recorded and contains the old workflowId: "5". But in the replay session, build-workflow created workflow "12". When the LLM tells the agent to run workflow "5", it doesn't exist.

The Solution: IdRemapper

The IdRemapper maintains a bidirectional mapping of old IDs to new IDs, learned incrementally as tools execute:

1. build-workflow executes → output: { workflowId: "12" }
2. IdRemapper compares recorded output { workflowId: "5" } with real output { workflowId: "12" }
3. Learns mapping: "5" → "12"
4. Next tool call: run-workflow({ workflowId: "5" })
5. IdRemapper translates input: run-workflow({ workflowId: "12" })
6. Tool executes successfully with the real ID

ID extraction is field-name aware — only fields named id or ending with Id (e.g., workflowId, executionId, credentialId) are compared. This prevents false mappings from unrelated data like execution output, web content, or file blobs.

Why the Proxy Can Ignore Request Bodies

During recording, the fixture's transform callback strips LLM request bodies down to an 80-character system prompt substring. This means MockServer matches requests by path and prompt prefix only, not by the full body. Since tool results flow into LLM request bodies (as tool_result content blocks), and those results now contain different IDs, the proxy would fail to match if it compared full bodies. By ignoring bodies, the proxy stays deterministic regardless of tool output content.

Shared State Across Runs

A single test may trigger multiple n8n "runs" — the orchestrator run, a background task follow-up, or a delegated sub-agent. The TraceIndex and IdRemapper are shared across all runs within one test (keyed by the test slug), so cursor positions and ID mappings persist correctly.

Two-Tier Tool Strategy

Tools are categorized by whether they can execute in CI:

Tier 1: Real Execution + ID Remapping (default)

Tools that only need the n8n database and engine. They execute for real, and the IdRemapper translates IDs in both directions.

The wrapping flow:

// Simplified — see langsmith-tracing.ts for full implementation
async execute(input, context) {
  const event = traceIndex.next(agentRole, toolName);  // Get recorded event
  const remappedInput = idRemapper.remapInput(input);   // Translate old IDs → new
  const realOutput = await tool.execute(remappedInput);  // Execute for real
  idRemapper.learn(event.output, realOutput);            // Discover new mappings
  return realOutput;                                      // Return real output
}

Tier 2: Pure Replay (external dependency tools)

Tools that need internet access or external services. They skip real execution entirely and return the recorded output (with ID remapping applied).

The wrapping flow:

async execute(input, context) {
  const event = traceIndex.next(agentRole, toolName);  // Validate tool sequence
  return idRemapper.remapOutput(event.output);          // Return recorded output
}

Not Wrapped

Some tools pass through without wrapping:

Trace Format

Each test's tool calls are recorded in trace.jsonl (newline-delimited JSON):

{"kind":"header","version":1,"testName":"should-approve-workflow-execution","recordedAt":"2026-04-09T12:00:00Z"}
{"stepId":1,"kind":"tool-call","agentRole":"orchestrator","toolName":"search-nodes","input":{...},"output":{...}}
{"stepId":2,"kind":"tool-call","agentRole":"workflow-builder","toolName":"build-workflow","input":{...},"output":{"workflowId":"5"}}
{"stepId":3,"kind":"tool-suspend","agentRole":"orchestrator","toolName":"run-workflow","input":{"workflowId":"5"},"output":{"denied":true},"suspendPayload":{...}}
{"stepId":4,"kind":"tool-resume","agentRole":"orchestrator","toolName":"run-workflow","input":{"workflowId":"5"},"output":{"executionId":"exec-100"}}

Event Types

• header — metadata (version, test name, timestamp)
• tool-call — normal tool invocation with input and output
• tool-suspend — human-in-the-loop tool paused for approval (includes suspend payload)
• tool-resume — tool resumed after user approval/denial (includes resume data)

TraceIndex

The TraceIndex groups events by agentRole with independent cursors per role. This handles interleaved orchestrator and sub-agent calls:

orchestrator: [search-nodes, run-workflow-suspend, run-workflow-resume]
                ^cursor=0
workflow-builder: [build-workflow]
                   ^cursor=0

When a tool is called, traceIndex.next(role, toolName) advances that role's cursor and validates the tool name matches. A mismatch means the agent diverged from the recorded path — the test fails with a clear error.

Proxy Expectations

Each test's HTTP exchanges are stored as individual JSON files:

expectations/instance-ai/should-send-message-and-receive-assistant-response/
  1775805992870-unknown-host-POST-_v1_messages-8a23f6c2.json    ← Anthropic API call
  1775805993100-api-staging.n8n.io-GET-_api_community_nodes-272f77d5.json  ← Node catalog
  trace.jsonl                                                     ← Tool trace

File Naming

<timestamp>-<host>-<method>-<path_slugified>-<8char_sha256>.json

• unknown-host = Anthropic API (CONNECT tunneling hides the real host)
• api-staging.n8n.io = n8n community nodes API

Sequential Loading

Expectations are loaded with sequential: true, which sets remainingTimes: 1 on each. They fire in order, one-shot. The last /v1/messages expectation is made unlimited: true to act as a fallback for any extra LLM calls caused by tool execution divergence.

Request Body Matching

LLM request bodies are replaced during recording with an 80-character substring of the system prompt. This is enough to distinguish between different types of calls (title generation vs. orchestrator vs. sub-agent) without being so specific that minor prompt changes break replay.

Test Infrastructure

Key Files

Test-Only Endpoints

Enabled by E2E_TESTS=true (set automatically by the Playwright fixture base):

Page Objects

• InstanceAiPage — chat input, send button, message selectors, preview iframe, artifact cards, confirmation panels
• InstanceAiSidebar — thread list, new thread button, thread-by-title lookup, rename input, action menu

Test Suites

Running Tests

Replay Mode (CI / no API key)

pnpm build:docker
pnpm --filter=n8n-playwright test:container:sqlite --grep "Instance AI"

Record Mode (local with real API key)

ANTHROPIC_API_KEY=sk-... pnpm --filter=n8n-playwright test:container:sqlite --grep "Instance AI"

This overwrites expectations/instance-ai/<test-slug>/ with fresh recordings.

Local-build mode (no docker, real Anthropic key)

For fast iteration against a local n8n build — skips the container and proxy stack entirely. Tests hit the real Anthropic API directly. This mode does not record proxy expectations.

cd packages/testing/playwright

pnpm test:local:instance-ai                  # full suite
pnpm test:local:instance-ai --grep "preview" # single test

Extra args flow through to playwright test:

# Single file
pnpm test:local:instance-ai instance-ai-workflow-preview.spec.ts

# Multiple instances in parallel — each gets its own random port + temp DB
pnpm test:local:instance-ai --grep "preview" &
pnpm test:local:instance-ai --grep "sidebar"  &
wait

# Pin the port (e.g. for browser inspection at http://localhost:5680)
N8N_BASE_URL=http://localhost:5680 pnpm test:local:instance-ai --grep "preview"

# Headed browser for visual debugging
pnpm test:local:instance-ai --grep "preview" --headed

test:local:instance-ai is a thin wrapper that pre-fills the instance-ai env vars over the generic test:local:isolated runner, which provides random free ports, a throwaway N8N_USER_FOLDER (so ~/.n8n is never touched), and process-group cleanup. See the [Playwright README](..