N8n Architect

n8n Architect

Use this skill for all n8n-as-code work: workspace readiness, environments, managed local instances, tunnels, workflow authoring, validation, sync, push, and pull.

Use {{N8NAC_CMD}} as the primary interface. Use {{N8N_MANAGER_CMD}} only for local managed runtime lifecycle, tunnels, and workflow presentation commands that are explicitly exposed by n8n-manager.

Context Root Protocol

• Treat the current context root as the directory containing n8nac-config.json, AGENTS.md, .agents/skills, and the configured workflowsPath.
• {{N8NAC_CONTEXT_ROOT_HINT}}
• Before any n8n work, first run {{N8NAC_CMD}} update-ai from the context root, then read AGENTS.md. update-ai is designed to create or refresh the n8n-as-code block without destroying existing user or agent instructions.
• Use the exact n8nac command and n8n-manager command listed in AGENTS.md. Those context-root commands override the portable examples in this skill.
• Run every {{N8NAC_CMD}} env ..., {{N8NAC_CMD}} workspace ..., {{N8NAC_CMD}} list, pull, push, validate, test, and update-ai command from the context root unless the user explicitly gives another context root.
• AGENTS.md is bootstrap context only, not a source of configuration truth.
• Do not infer environment, project, or workflowsPath from AGENTS.md.
• Before n8n work, resolve the effective context from the backend:

bashCopy code
{{N8NAC_CMD}} env status --json

• Use the returned workflowsPath for workflow files. It is the configured workflow directory for the active environment.
• Do not reconstruct workflowsPath from environment name/id, instanceIdentifier, instanceUserIdentifier, projectId, projectName, or legacy sync fields.
• Never write n8nac-config.json, ~/.n8n-manager, or n8n-manager secret files by hand.

Workspace Readiness

Resolve the effective environment through the backend before workflow work:

bashCopy code
{{N8NAC_CMD}} env status --json

• Treat env status --json as the source of effective workspace readiness.
• Do not infer readiness from raw files, generated agent docs, or directory names.
• A workspace .env holding N8N_HOST (plus any of N8N_API_KEY, N8N_NATIVE_MCP_URL, N8N_NATIVE_MCP_TOKEN) configures the default environment on first use: env status --json then resolves with no env add, env auth set, env use, or native-mcp configure.
• If env status --json fails because the workspace is not configured, use env add, env auth set, and env use to create or select a V4 workspace environment.
• Managed local instances remain machine-global runtime resources.
• Workspace environments remain workspace-scoped and are managed through {{N8NAC_CMD}} env ....

Bootstrap Order

1. cd to the context root.
2. Run {{N8NAC_CMD}} update-ai, then read AGENTS.md.
3. Run {{N8NAC_CMD}} env status --json. If it resolves with accessStatus: "ready", the workspace is ready — skip steps 4-9. A resolution with missing-api-key or invalid-api-key means only the credential is missing: go to step 7.
4. If the context root is not ready, inspect managed local instances with {{N8N_MANAGER_CMD}} instance list.
5. Reuse an existing environment or managed local instance when suitable.
6. If no suitable environment exists, stop and ask the user whether they want to connect a remote n8n URL or create/reuse a managed local n8n instance. Do not create infrastructure by default. If the user chooses a managed local instance, ask separately whether they want a public tunnel.
7. Ask for host/API key only for an explicitly remote n8n environment.
8. Configure the environment with:

bashCopy code
{{N8NAC_CMD}} env add <name> --base-url <url> --workflows-path workflows/<name>
{{N8NAC_CMD}} env auth set <name> --api-key-stdin
{{N8NAC_CMD}} env use <name>

For a managed local instance:

bashCopy code
{{N8NAC_CMD}} env add Local --managed-instance <id> --workflows-path workflows/local
{{N8NAC_CMD}} env use Local

9. Run {{N8NAC_CMD}} update-ai after changing environments when the facade does not do it automatically.

Environments

Use {{N8NAC_CMD}} env ... for workspace environments, remote URLs, active environment, API-key binding, projects, and workflow paths.

bashCopy code
{{N8NAC_CMD}} env status --json
{{N8NAC_CMD}} env list
{{N8NAC_CMD}} env add <name> --base-url <url> --workflows-path workflows/<name>
{{N8NAC_CMD}} env auth set <name> --api-key-stdin
{{N8NAC_CMD}} env use <name>

• Prefer --api-key-stdin for API keys.
• Prefer env add --pin to create and pin the default environment in one process instead of a separate env use.
• Do not pass secrets inline in shell arguments.
• env auth set binds the key to one environment, so several environments may share a base URL with one key each. Run it once per environment; apiKeySource in env status --json is workspace-environment when the environment uses its own key.
• Do not ask for host/API key when the user wants a managed local Docker instance.
• Do not print API keys or credential secret values back to the user.
• If a command or flag is unfamiliar, run {{N8NAC_CMD}} env --help or {{N8NAC_CMD}} env <subcommand> --help.

Attach a managed local instance to the workspace with {{N8NAC_CMD}} env ...:

bashCopy code
{{N8NAC_CMD}} env add Local --managed-instance <id> --workflows-path workflows/local
{{N8NAC_CMD}} env use Local

Managed Local Runtime

Use {{N8N_MANAGER_CMD}} only for local managed instance lifecycle, tunnels, and workflow presentation commands that are part of the local runtime layer.

Inspect existing managed instances before changing local machine state:

bashCopy code
{{N8N_MANAGER_CMD}} instance list
{{N8N_MANAGER_CMD}} instance --help
{{N8N_MANAGER_CMD}} config get

Do not invent n8n-manager subcommands. Use {{N8N_MANAGER_CMD}} <subcommand> --help when unsure.

When the context root is not configured and no suitable existing instance is available, stop and ask the user to choose. Do not create infrastructure by default.

Present these choices clearly:

• use an existing managed local instance if one is available;
• create a new managed local n8n instance;
• configure a remote n8n URL as a workspace environment through {{N8NAC_CMD}} env.

If the user chooses a managed local Docker instance, ask the tunnel question separately:

• without public tunnel: local n8n only, suitable for normal UI/API workflow work;
• with public tunnel: exposes the instance through a public URL, useful for webhooks/forms/chat triggers and remote callbacks.

Do not enable, refresh, or start a public tunnel unless the user explicitly requested public access, webhook testing, or approved the tunnel option. If public access is not needed, create/start the managed instance without --tunnel.

Only run these commands after the user has explicitly chosen the corresponding option.

Managed local instance without public tunnel:

bashCopy code
{{N8N_MANAGER_CMD}} instance create
{{N8N_MANAGER_CMD}} instance start <id>
{{N8N_MANAGER_CMD}} instance list

Managed local instance with public tunnel:

bashCopy code
{{N8N_MANAGER_CMD}} instance create
{{N8N_MANAGER_CMD}} instance start <id>
{{N8N_MANAGER_CMD}} tunnel start <id>

Instance and tunnel operations are per managed local instance:

bashCopy code
{{N8N_MANAGER_CMD}} instance start <id>
{{N8N_MANAGER_CMD}} instance stop <id>
{{N8N_MANAGER_CMD}} instance remove <id>
{{N8N_MANAGER_CMD}} tunnel start <id>
{{N8N_MANAGER_CMD}} tunnel stop <id>

• Do not delete local instance data unless the user explicitly asks for destructive deletion.
• If Docker is unavailable or the daemon is stopped, report the backend diagnostic and stop. Do not loop.
• If a command fails repeatedly, stop after two attempts and explain the backend diagnostic.

Sync Discipline

• Pull before reading or modifying an existing workflow.
• Push after every modification.
• Use list to inspect workflow IDs, file paths, and sync status.

bashCopy code
{{N8NAC_CMD}} list
{{N8NAC_CMD}} pull <workflowId>
{{N8NAC_CMD}} push <path-to-workflow.workflow.ts> --verify

• push requires the full workflow file path, either absolute or context-root-relative. Do not pass a bare filename.
• On n8n 2.x, pushing to a published workflow also releases it to production — the API re-publishes on update. Treat every push to a published workflow as a deploy. Use push --draft when the user wants to check the change in n8n first: it re-pins the previously published version so production keeps running what it already ran.
• For a new workflow, create the file inside the workflowsPath returned by env status --json, then confirm it with {{N8NAC_CMD}} list --local.
• If push/pull reports a conflict, use explicit resolution commands. Do not overwrite remote changes blindly.
• pull and conflict resolution operate on a single workflow ID.
• list is the lightweight command that covers all workflows at once.
• If you skip pull, a later push can be rejected by optimistic concurrency control when the remote changed.

Conflict Handling

If push or pull reports a conflict, stop and inspect the conflict. Use explicit resolution commands only after choosing the intended direction:

bashCopy code
{{N8NAC_CMD}} resolve <workflowId> --mode keep-current
{{N8NAC_CMD}} resolve <workflowId> --mode keep-incoming

• keep-current force-pushes the local version.
• keep-incoming force-pulls the remote version.
• Never silently force-push over a remote change.

Schema-First Research

Never guess n8n node parameters.

bashCopy code
{{N8NAC_SKILLS_CMD}} examples search "<workflow pattern>"
{{N8NAC_SKILLS_CMD}} search "<node or capability>"
{{N8NAC_SKILLS_CMD}} node-info <nodeName>
{{N8NAC_SKILLS_CMD}} validate <workflow.workflow.ts>

• Batch node queries: query multiple nodes in a single command using {{N8NAC_SKILLS_CMD}} node-info <node1> <node2> ... --compact to inspect essential properties, required parameters, and valid options in one fast, token-efficient turn.
• Use exact node type and valid typeVersion values from node-info.
• Use exact resource, operation, option, and parameter names from schema output.
• Parameters of type resourceLocator must be structured as { __rl: true, value: '...', mode: 'list' | 'id' | 'url' }.
• Do not invent parameters, operations, credential types, or CLI flags.
• Treat schema output as the absolute source of truth even if examples or memory disagree.
• Prefer the highest valid typeVersion returned by schema output.
• For fixed collections such as Switch/If rules, Wait form fields, or nested options, read the full node-info output before writing values.

Optional Native n8n MCP Assist

The n8n-as-code MCP server is a client adapter for N8NAC tools. The native n8n MCP server is a separate live n8n instance endpoint. Native n8n MCP can complement this workflow for native knowledge, live state, and runtime execution, but it does not replace {{N8NAC_CMD}}, bundled knowledge, .workflow.ts, Git, or the sync discipline.

Use this routing policy:

• Default to local {{N8NAC_CMD}} for code-first workflow authoring, validation, pull, push, credentials, execution history, and presentation. Use {{N8NAC_SKILLS_CMD}} as the bundled offline knowledge default.
• Native MCP assist is configured per n8n-as-code environment. When creating or updating an environment, offer to configure it with {{N8NAC_CMD}} native-mcp configure <environment> --token-stdin; do not ask the user to manually configure a separate MCP server for Claude Code or the VS Code Workbench.
• If native MCP assist is configured, use it where it complements n8n-as-code: read-only live discovery, server-side validation, native SDK/reference knowledge, live node definitions, credential metadata without secrets, execution inspection, projects, folders, and explicit runtime execution/test strategy when supported.
• Check native availability with {{N8NAC_CMD}} native-mcp status --include-tools --json before relying on native tools.
• For user requests about the current/live n8n instance, existing remote workflows, available nodes in this instance, credential metadata, projects, folders, executions, drift, or duplicate discovery, prefer native MCP read-only tools after the status check. Do not fall back to local {{N8NAC_CMD}} list, fetch, verify, or bundled skills as the primary source for those live-audit facts when native MCP read-only tools are available.
• Do not expose native MCP assist on non-loopback HTTP/SSE transports unless the MCP transport is authenticated and N8NAC_NATIVE_MCP_ALLOW_REMOTE=1 is explicitly set.
• Do not request full live execution payloads with includeData=true unless the user explicitly needs payload data and N8NAC_NATIVE_MCP_ALLOW_EXECUTION_DATA=1 is set.
• Prefer {{N8NAC_CMD}} test when the execution strategy is to exercise the real webhook, chat, or form trigger contract.
• Prefer native runtime execution only when the generated execution strategy explicitly calls for it and it does something better than {{N8NAC_CMD}} test, such as workflow ID execution, non-webhook workflow testing, native pin-data test preparation, or direct execution diagnostics.
• Treat native execute/test as a side-effecting runtime action, like {{N8NAC_CMD}} test; do not run it just because the tool exists.
• Do not use native MCP create, update, publish, unpublish, archive, or destructive data-table tools unless the user explicitly requests direct native MCP mode and the tool is gated by permissions.
• If a workflow is ever created or changed through native MCP direct mode, immediately pull it back with {{N8NAC_CMD}} pull <workflowId> so the .workflow.ts file and Git remain the source of truth.
• If native MCP validation and local validation disagree, stop and report the divergence instead of forcing a push or direct update.
• Never put native MCP tokens in project files, generated docs, command arguments, or responses.

Use-case routing examples:

• Workflow authoring, editing, pull, push, sync, credentials, and durable workflow changes: use local {{N8NAC_CMD}} commands and .workflow.ts files.
• Offline node knowledge, examples, documentation, and schema-first authoring: use local {{N8NAC_SKILLS_CMD}} commands first.
• Live workflow discovery, drift investigation, projects, folders, credentials metadata, duplicate discovery, and execution inspection: prefer native MCP read-only tools when configured because the user is asking for current instance state.
• Connected-version node definitions or server-side validation: prefer native MCP read-only tools when the user asks what is available in this instance or needs validation against the connected n8n version. Use bundled knowledge for offline authoring when live instance state is not needed.
• Runtime execution: prefer {{N8NAC_CMD}} test for real webhook, chat, or form trigger contracts; prefer native runtime execution only for explicit workflow-ID execution, non-webhook testing, native pin-data preparation, or direct execution diagnostics.
• Direct native workflow creation, update, publish, unpublish, archive, or destructive operations: do not use them as an automatic path; require an explicit direct-native request and sync-back plan.

Do not treat the presence of any MCP server as permission to call native n8n MCP tools. Native n8n MCP is used if and only if the generated execution or investigation strategy needs live n8n capabilities that local N8NAC cannot provide as well.

Native MCP assist is a complementary knowledge, live-state, and runtime enrichment path, not the primary authoring or sync path.

Knowledge Commands

If your runtime exposes n8n-as-code MCP tools (search_n8n_knowledge, get_n8n_node_info, search_n8n_workflow_examples, validate_n8n_workflow), prefer them over the commands below. They answer from a resident process that parses the ontology once, so every call after the first is effectively free, while each shell command pays a fresh process start.

Otherwise, use these commands instead of guessing:

bashCopy code
{{N8NAC_SKILLS_CMD}} search "<node or capability>"
{{N8NAC_SKILLS_CMD}} node-info <nodeName>
{{N8NAC_SKILLS_CMD}} node-schema <nodeName>
{{N8NAC_SKILLS_CMD}} docs "<topic>"
{{N8NAC_SKILLS_CMD}} guides "<topic>"
{{N8NAC_SKILLS_CMD}} examples search "<workflow pattern>"
{{N8NAC_SKILLS_CMD}} examples info <id>
{{N8NAC_SKILLS_CMD}} examples download <id>

• Prefer --compact on search, node-info, and node-schema: same schemas, bounded output (required params + snippet + gating flags).
• Prefer one batch --compact over N separate lookups: one process parses the ontology once. Pass --calls '<json>', --calls-file <path> (file avoids shell-quoting), or pipe JSON via stdin. --compact applies to search, node-info, node-schema; examples-search and examples-info always return full workflow data. Example: {{N8NAC_SKILLS_CMD}} batch --compact --calls '[{"cmd":"search","query":"gmail"},{"cmd":"node-info","name":"gmailTool"}]'.
• For several nodes at once, node-info and node-schema also take multiple names directly: {{N8NAC_SKILLS_CMD}} node-info <node1> <node2> ... --compact.
• Start with examples search when the user asks for a common automation pattern.
• Fetch community examples only when you do not know how to wire something, when the workflow is unusually complex, or when the user explicitly asks. Each download costs a full roundtrip: for routine tasks, local knowledge (search, node-info, batch) is faster and authoritative. Skip examples otherwise.
• Use examples to learn patterns, not as authority over current node schemas.
• If a command or flag is unfamiliar, run {{N8NAC_CMD}} <subcommand> --help; do not invent flags.

Workflow Authoring Rules

• Use TypeScript decorators from @n8n-as-code/transformer.
• Regular nodes connect with source.out(0).to(target.in(0)).
• AI sub-nodes connect with .uses(), never .out().to().
• ai_tool and ai_document connections are arrays: ai_tool: [this.Tool.output].
• Other AI connection types are single refs, such as ai_languageModel: this.Model.output.
• They also accept an array when a node exposes several inputs of the same type, where the position is the input index: ai_languageModel: [this.Model.output, this.FallbackModel.output] (fallback model, Model Selector).
• Check node-info for connection-dependent boolean flags before declaring .uses() connections.

Every .workflow.ts file starts with a <workflow-map> block. Read that map first, locate the property name you need, then read only the relevant class section.

Minimal Workflow Structure

typescriptCopy code
import { workflow, node, links } from '@n8n-as-code/transformer';

@workflow({
  name: 'Workflow Name',
  active: false
})
export class MyWorkflow {
  @node({
    name: 'Descriptive Name',
    type: '/* exact type from node-info */',
    version: 4,
    position: [250, 300]
  })
  MyNode = {
    /* parameters from node-info */
  };

  @node({
    name: 'Next Node',
    type: '/* exact type from node-info */',
    version: 3,
    position: [520, 300]
  })
  NextNode = {};

  @links()
  defineRouting() {
    this.MyNode.out(0).to(this.NextNode.in(0));
  }
}

Expression Syntax

• Prefer modern expressions: {{ $json.fieldName }}.
• Use specific-node expressions when needed: {{ $('Node Name').item.json.field }}.
• Avoid legacy $node["Name"].json.field unless you are preserving an existing workflow and have a reason.
• In Switch/If comparisons, value1 is the expression being evaluated and value2 is the literal comparison value.

Node Naming

• Use descriptive names such as Get Customers, Send Slack Alert, or Normalize Payload.
• Avoid names like Node1, HTTP Request, or Code when a more specific name is available.
• Connection references must match the exact node property names in the TypeScript class.

Reading Workflow Files Efficiently

Use the <workflow-map> block as the index before loading large workflow files.

typescriptCopy code
// <workflow-map>
// Workflow : My Workflow
// Nodes   : 12  |  Connections: 14
//
// NODE INDEX
// Property name                    Node type (short)         Flags
// ScheduleTrigger                  scheduleTrigger
// AgentGenerateApplication         agent                      [AI] [creds]
// OpenaiChatModel                  lmChatOpenAi               [creds] [ai_languageModel]
// Memory                           memoryBufferWindow         [ai_memory]
// GithubCheckBranchRef             httpRequest                [onError->out(1)]
//
// ROUTING MAP
// ScheduleTrigger
//   -> Configuration
//     -> BuildProfileSources -> LoopOverProfileSources
//
// AI CONNECTIONS
// AgentGenerateApplication.uses({ ai_languageModel: OpenaiChatModel, ai_memory: Memory })
// </workflow-map>

Navigation rule:

1. Read <workflow-map> first.
2. Locate the property name you need.
3. Search for that property in the file.
4. Read only the relevant node or routing section unless broader context is required.

AI And LangChain Node Rules

AI sub-nodes are not regular data-flow nodes.

typescriptCopy code
@links()
defineRouting() {
  this.ChatTrigger.out(0).to(this.AiAgent.in(0));

  this.AiAgent.uses({
    ai_languageModel: this.OpenaiModel.output,
    ai_memory: this.Memory.output,
    ai_outputParser: this.OutputParser.output,
    ai_tool: [this.SearchTool.output],
  });
}

• Use .uses() for language models, memory, tools, parsers, embeddings, vector stores, retrievers, and other AI sub-nodes.
• Never connect AI sub-nodes with .out().to().
• ai_tool and ai_document must be arrays; every entry lands on input index 0.
• Most other AI connection types are single refs, or an array when the node exposes several inputs of the same type — position = input index.
• needsFallback: true (Agent, Basic LLM Chain) needs a second model on input 1: ai_languageModel: [this.Model.output, this.FallbackModel.output]. Same for the Model Selector node.
• Some nodes require boolean flags to expose AI ports or gated parameters. Check node-info before declaring .uses().

Common Mistakes To Avoid

• Wrong node type: use the exact full type returned by schema output, including package prefix when provided.
• Outdated or non-existent typeVersion: use a value from the schema output.
• Invalid operation/resource value: use exact option values from the schema.
• Mismatched resource and operation: each resource enables its own operations.
• Guessing nested structures: fixed collections have exact shapes.
• Wrong connection names: match TypeScript property names exactly.
• Inventing nodes, credentials, operations, or parameters.
• Connecting AI sub-nodes with .out().to().
• Using ai_tool: this.Tool.output instead of ai_tool: [this.Tool.output].
• Inverting Switch/If value1 and value2.
• Using old Wait form structures such as formFieldsUi.fieldItems when the current schema expects formFields: { values: [...] }.
• Passing a bare filename to push.
• Treating Class A runtime/config gaps as workflow-code bugs.

Verify, Test, And Present

Prefer push --verify: it fetches the pushed workflow and validates it in the same process. A standalone verify right after push --verify re-checks the same state — skip it unless you pushed without --verify.

bashCopy code
{{N8NAC_CMD}} push <path> --verify
{{N8NAC_CMD}} test-plan <workflowId> --json

For webhook, chat, or form workflows, prefer the production test sequence:

bashCopy code
{{N8NAC_CMD}} workflow activate <workflowId>
{{N8NAC_CMD}} test <workflowId> --prod

• Class A configuration gaps require user/config action, not workflow rewrites.
• Runtime-state issues such as unarmed test webhooks are not workflow-code bugs.
• Class B wiring errors are fixable in the workflow file.
• Stop after two repeated failures with the same diagnostic.

Workflow Presentation Contract

{{N8NAC_CMD}} workflow present is the standard way to show a workflow to the user. It is v4-environment aware and part of the workflow authoring loop.

Run it whenever one of these is true:

• you created a workflow;
• you modified and pushed a workflow;
• you ran or tested a workflow and the user needs to inspect it;
• the user asks to show, open, present, display, or give the URL/link for a workflow.

bashCopy code
{{N8NAC_CMD}} workflow present <workflowId> --json

Rules:

• Do not manually construct n8n workflow URLs.
• Do not return an internal local n8n URL when a presentation URL is available.
• Use the url returned by workflow present --json as the user-facing URL.
• If you do not know the workflow ID, run {{N8NAC_CMD}} list first and select the matching workflow.
• Do not call {{N8N_MANAGER_CMD}} presentWorkflowResult; it is a legacy runtime command and is not workspace-environment aware.
• If workflow present fails, report the backend diagnostic and then provide the best direct n8n URL only as a fallback.
• Do this before the final response when the task created, changed, pushed, ran, or explicitly asks to show a workflow.

Testability Protocol

For webhook, chat, or form workflows:

1. Push with verification when possible.
2. Run test-plan to inspect trigger type, endpoint, and suggested payload.
3. Activate the workflow.
4. Test with --prod by default.

bashCopy code
{{N8NAC_CMD}} push <path-to-workflow.workflow.ts> --verify
{{N8NAC_CMD}} test-plan <workflowId> --json
{{N8NAC_CMD}} workflow activate <workflowId>
{{N8NAC_CMD}} test <workflowId> --prod

Use bare {{N8NAC_CMD}} test <workflowId> only when a test URL was intentionally armed in the n8n editor.

For GET/HEAD webhooks that read from $json.query, prefer:

bashCopy code
{{N8NAC_CMD}} test <workflowId> --query '{"key":"value"}' --prod

Execution Debugging

If a webhook returns success but the workflow behavior is wrong, inspect executions instead of guessing:

bashCopy code
{{N8NAC_CMD}} execution list --workflow-id <workflowId> --limit 5 --json
{{N8NAC_CMD}} execution get <executionId> --include-data --json

• A successful HTTP trigger only means n8n accepted the request.
• The execution can still fail later inside the workflow.
• Use execution data to identify the failing node and real payload shape.

Credential Workflow

When a workflow is blocked by missing credentials, resolve the credential gap without rewriting unrelated workflow logic.

bashCopy code
{{N8NAC_CMD}} workflow credential-required <workflowId> --json
{{N8NAC_CMD}} credential schema <type>
{{N8NAC_CMD}} credential list --json
{{N8NAC_CMD}} credential create --type <type> --name <name> --file cred.json --json
{{N8NAC_CMD}} workflow activate <workflowId>

• workflow credential-required exits non-zero when at least one credential is missing. Treat that as a signal to act, not as a workflow-code failure.
• Use credential schema to discover required fields.
• Ask the user for secret values when needed.
• Prefer --file for credential creation. Do not pass secrets inline in shell arguments.
• Do not print API keys or credential secret values back to the user.
• If credential creation fails, read the validation message and change the payload before retrying.

Operating Loop

For most workflow tasks:

1. Resolve context with env status --json.
2. Read workflowsPath from the backend response.
3. Inspect existing workflows with list.
4. Pull before editing an existing workflow.
5. Search examples and schemas.
6. Edit or create the .workflow.ts file.
7. Validate locally.
8. Push with --verify.
9. Test if the workflow is HTTP-triggered.
10. Inspect executions when behavior is unclear.
11. Present the final workflow link with {{N8NAC_CMD}} workflow present <workflowId> --json.

Response Discipline

• Explain concrete actions and command results, not generic capability.
• When the user asks for an URL or visual inspection of a workflow, run {{N8NAC_CMD}} workflow present <workflowId> --json instead of composing a URL manually.
• If setup is missing, use n8nac env ... for workspace environments and n8n-manager only for managed local instances.
• Do not ask for host/API key unless the user chooses a remote n8n environment.
• Do not tell the user to run setup commands when you can run non-interactive commands yourself.
• Stop after two repeated failures with the same diagnostic and report the backend error clearly.