N8n Workflow Lifecycle Official

Workflow review checklist

Severity-tiered audit for any existing n8n workflow or n8n build (group of workflows). Different from VALIDATION_CHECKLIST.md (which is pre-publish gates for your own in-progress build): this file is for reviewing arbitrary workflows, including ones built by anyone, from anywhere.

How to use

Walk the list top to bottom. For each item, inspect the workflow (get_workflow_details) and decide if the issue applies. Report findings grouped by severity. Each item links to the canonical skill reference for the why and the how to fix. Be a very thourough reviewer and air on the side of reading references to ensure full context.

You're reviewing JSON, not SDK source. get_workflow_details returns the n8n workflow JSON (nodes with parameters, connections graph, credential references, node type strings like n8n-nodes-base.httpRequest). Phrase findings in JSON terms: "node Foo has parameters.onError set to 'continueErrorOutput' but connections.Foo.main[1] is empty,".

Cross-cutting first

Before walking the per-domain list:

• [ ] Pull the workflow(s). get_workflow_details({ workflowId }), required so subsequent checks operate on the actual JSON, not assumptions.

• [ ] High-level intent / logic smell test. Read the workflow's description (if it exists), then trace the happy path once top to bottom. Does the structure match what the description says it does? Anything obviously dead, missing, contradictory, or not fitting (a write node in a workflow described as read-only, a fan-out with one terminal branch that should be wired further, an HTTP call to a domain unrelated to the stated integration)? Catches whole classes of issues the per-domain checks won't surface.

• [ ] Note the trigger type. Webhook, schedule, manual, sub-workflow, chat trigger. Severity of issues changes by trigger (a webhook-API workflow needs error paths; a manual run does not).

• [ ] Note whether the workflow is active. Active/published workflows with broken connections are higher severity.

MUST FIX

Credentials and secrets

• [ ] Tokens, API keys, or secrets in node text fields. Any node parameter holding Bearer xxx, sk-..., an API key, or a password as plain text. The credential system is the only correct home. → n8n-credentials-and-security-official

• [ ] Tokens stored in Set node values for later {{$json.token}} referencing. The token is in workflow JSON regardless of how it's read. → n8n-credentials-and-security-official

• [ ] Hardcoded credentials in Code nodes. Same leak surface as text fields. → n8n-code-nodes-official anti-patterns

• [ ] HTTP Request nodes with Authorization header values typed in directly instead of using a credential. For Authorization: Bearer <token>, use Bearer Auth (httpBearerAuth) so the token is stored without the prefix. For other custom auth headers, use Header Auth (httpHeaderAuth). → HTTP_REQUEST_WITH_AUTH.md

• [ ] Secret read from $vars.X and used as an auth value. Use a credential instead. → n8n-credentials-and-security-official

SQL / query injection

• [ ] User input interpolated into a query string. Any DB node with parameters.query containing {{ $json.something }} (or any {{ ... }} expression that resolves from caller input) inside the SQL itself is a SQL injection: n8n substitutes the expression into the query before parameter binding. Use parameter binding instead (Postgres / MySQL: $1, $2 placeholders + parameters.options.queryReplacement; Mongo: object filters). → DATABASE_NODES.md

Connection bugs (silent breakage)

• [ ] Merge index off-by-one. parameters.useDataOfInput is 1-indexed but the corresponding entry in connections.<source>.main[index] is 0-indexed. If the merge node's expected primary input doesn't match the wiring, the wrong source is picked silently. → MERGE_NODE.md

• [ ] Merge with 3+ sources but numberOfInputs left at default 2. Third source silently drops. → MERGE_NODE.md

• [ ] Error output wired without onError: 'continueErrorOutput' on the node config. Error branch is unreachable; node failure halts the workflow. → NODE_ERROR_OUTPUTS.md

• [ ] onError: 'continueErrorOutput' set but main[1] not wired. Error path is enabled but goes nowhere. → NODE_ERROR_OUTPUTS.md

Webhook API workflows (Webhook + Respond to Webhook)

• [ ] Webhook performs a sensitive action with parameters.authentication: 'none'. "Sensitive" = mutates state, sends external messages, hits production data, exposes private info, triggers paid actions. Anyone with the URL can fire it. Set parameters.authentication to 'basicAuth' or 'headerAuth' and use the matching credential. → n8n-credentials-and-security-official

Sub-workflow contracts

• [ ] Execute Workflow Trigger set to passthrough when it shouldn't be. Passthrough loses the typed-input contract that agent tools (fromAi()) and structured callers need. Only correct when (a) the sub-workflow specifically receives binary AND isn't an agent tool, or (b) the sub-workflow takes no inputs (Define Below requires at least one field). For (b), the body should open with a Set ("Keep Only Set", no fields) and the trigger should carry a sticky noting no inputs are expected. → n8n-subworkflows-official non-negotiables and SUBWORKFLOW_PATTERNS.md "Splitting by input shape"

First-class n8n Agents

Review an Agent with get_agent (draft) and get_agent with versionId set to its published activeVersionId (live). Findings are in config JSON terms.

• [ ] No memory block on a conversational Agent. It forgets the previous turn even inside one session. Add { "enabled": true, "storage": "n8n" }. → MEMORY.md

• [ ] Saved sub-agent unpublished while the parent is published. Production delegation fails and the parent answers without it. Nothing else flags this. → SUB_AGENTS.md

• [ ] Tool that sends, pays, deletes, or changes data without requireApproval (or MCP approval). → APPROVALS.md

• [ ] Approval-required tool on an Agent that runs from Message an Agent or a scheduled task. Fails, or parks with no one to answer. → APPROVALS.md

• [ ] Telegram channel with accessMode: public when the Agent isn't meant for strangers. Anyone can talk to it and spend its budget. → CHANNELS.md

• [ ] Two different Agents called with the same Message an Agent session key in one project (for example both keyed on a bare Telegram id). Keys are project-wide, not per Agent, so the second Agent's call fails with "Failed to create the agent execution record". Prefix keys per Agent. → MESSAGE_AN_AGENT.md

Chat-triggered agents (Slack / Discord / Teams / Telegram)

• [ ] Bot's own user ID not filtered out, either via the trigger's own filter option (preferred: Slack's options.userIds exclusion list) or as the first node after the trigger. The bot's reply re-triggers the workflow → infinite loop. Watch out for surface-specific semantics: Telegram's userIds is an allowlist, not an exclusion list. → CHAT_SURFACE_PATTERNS.md

SHOULD FIX

Naming

• [ ] Generic node names (HTTP Request1, Set2, Postgres1). Debugging-hostile: a failure on node "HTTP Request3" tells the operator nothing, but node "Fetch order details" localizes the break instantly. Rename every node to describe what it does in this workflow. → NAMING_CONVENTIONS.md

Comms nodes (Slack, Gmail, Discord, SMTP, Telegram)

• [ ] n8n attribution still appended on Slack / Gmail / Email / Discord nodes. Most comms nodes have "Append n8n Attribution" enabled by default. Users typically want it removed in production. → COMMS_NODES.md

• [ ] Slack thread reply posting as top-level message. thread_ts not set, or set in the wrong nested location. → COMMS_NODES.md

Set node antipattern

• [ ] Set node feeding only 0 or 1 downstream consumer. Most common antipattern in the pack. Delete + inline the expression at the consumer. → n8n-expressions-official "The Set-node antipattern"

• [ ] Set node before a Data Table Insert/Update mapping fields to schema. Map directly in the Data Table node's per-column expression slots. → n8n-data-tables-official strong defaults

• [ ] Set node building an email/Slack body. Build the body inline in the comms node's body field with an expression. → COMMS_NODES.md

• [ ] Multiple consecutive Set nodes each defining one field. Collapse, or eliminate. → n8n-expressions-official

Code node antipattern

• [ ] Code node doing pure data shaping (.map, .filter, .find, field rename, optional chaining). Use an expression or Edit Fields with arrow function. → n8n-code-nodes-official decision tree

• [ ] Code node using crypto.createHash / crypto.createHmac. Use the native Crypto node (n8n-nodes-base.crypto). Recurring AI slip. → n8n-code-nodes-official "Cryptographic operations"

• [ ] Code node parsing XML / SOAP / RSS. Use the native XML node + Edit Fields with arrow function for extraction. → n8n-code-nodes-official "XML / SOAP / RSS parsing"

• [ ] Code node + Set node combo (Set builds inputs, Code transforms). One Edit Fields with arrow function does both. → ARROW_FUNCTIONS_IN_EDIT_FIELDS.md

Expression discipline

• [ ] $json.x references deep in workflows with branches/intermediates. Switch to $('Source Node').item.json.x for refactor stability. → n8n-expressions-official non-negotiable

• [ ] DateTime nodes used for date math/formatting. Use Luxon expressions (DateTime.fromISO(...)) inline. → n8n-expressions-official strong defaults

• [ ] $env.X referenced in any expression. Doesn't work, throws at runtime. Replace with $vars.X (paid plans), a Data Table, or a credential for secrets. → n8n-expressions-official anti-patterns

• [ ] Aggregate node + per-item execution mismatch. Expressions using $input.all() / $('Node').all() without combining with another node's .item should set executeOnce: true on the node. → n8n-loops-official non-negotiable and n8n-expressions-official ".all().map() triggers an executeOnce question"

Execution model

• [ ] Workflow assumes fan-out branches execute in parallel. They don't, n8n runs them sequentially top-to-bottom by Y-position. Real concurrency needs sub-workflow dispatch with mode: 'each' + waitForSubWorkflow: false. → n8n-workflow-lifecycle-official "Execution model"

Loops

• [ ] Loop Over Items added "to make it loop" when default per-item iteration handles it. Default per-item iteration already waits for each item before the next, so a Loop Over Items added "to wait for all items" is unnecessary. → n8n-loops-official "When NOT to reach for Loop Over Items"

• [ ] Custom pagination implementation (Loop Over Items + $pageCount, hand-rolled while in a Code node, Set + IF cycle, etc.) instead of HTTP Request's built-in Pagination option. → HTTP_PAGINATION.md

• [ ] Reset-mode loop with no clear termination. reset: true without an explicit stop condition + $runIndex ceiling = infinite loop, n8n eats memory until killed. → LOOP_OVER_ITEMS.md "Reset mode"

• [ ] One Loop Over Items nested inside another in the same workflow. Doesn't work; breaks at runtime. Move the inner loop into a sub-workflow called per outer iteration (mode: 'each'). → LOOP_OVER_ITEMS.md "Nesting Loop Over Items"

Self-healing on transient failures

• [ ] Network-calling nodes (HTTP, comms, DB, AI) without retryOnFail configured. Transient 429s and upstream blips surface as 5xx, alerts fire on noise. → n8n-error-handling-official "Self-healing on transient failures"

Switch nodes

• [ ] Switch with no fallback output configured. Unmatched items silently drop. Set options.fallbackOutput: 'extra' and options.renameFallbackOutput: '<name>'. → SWITCH_FALLBACK.md

• [ ] Switch outputs unnamed. Set renameOutput: true + outputKey: '<name>' per rule for self-documenting branches. → SWITCH_FALLBACK.md

Sub-workflows

• [ ] Duplicated logic across workflows that would be a sub-workflow. → n8n-subworkflows-official decision tree

• [ ] Sub-workflow with no description. Won't be found in future searches; nobody (or AI) knows what it does. → n8n-subworkflows-official anti-patterns

• [ ] Sub-workflow has hidden side effects. The name and description describe pure logic (parse, validate, format, compute, transform), but the body contains write / send nodes (Data Table Insert/Update, Slack/Gmail/Discord send, HTTP POST, file write, audit log, etc.). Callers reasonably assume the sub-workflow is safe to retry; doing so creates duplicate writes or sends. Either declare the side effect (rename to e.g. Audit: or <Domain>:, document it in description, return a result the caller can branch on) or move the side effect out of this sub-workflow. → n8n-subworkflows-official "Stateless vs stateful"

• [ ] ~30-node workflow with no extraction. Extract logical sections into sub-workflows. → n8n-subworkflows-official

AI Agents and tools

• [ ] options.maxIterations left at default 10 on a multi-tool agent. Likely too low for modern agents with flexible tool sets; throws "Max iterations reached" workflow error. Raise to 30-50+. → AI_NODES.md "Iteration cap"

• [ ] Generic tool names (doStuff, runQuery). Model can't tell which tool to pick, skips them or hallucinates parameters. Use verb-first specific names. → TOOLS.md

• [ ] Default, Empty, or one-line tool descriptions. Model has no clue when to invoke. Tool descriptions are part of the prompt. → TOOLS.md

• [ ] outputParserStructured without autoFix: true. One bad model output and the workflow fails. Set autoFix: true with a coding-capable fixer model. → STRUCTURED_OUTPUT.md

• [ ] Tools with user-visible side effects (send, pay, refund) without human review. Wrap with slackHitlTool / discordHitlTool / telegramHitlTool / gmailHitlTool / etc. → HUMAN_REVIEW.md

• [ ] Approval message via fromAi() instead of $tool.parameters.<name>. Model paraphrases; you approve text not values. → HUMAN_REVIEW.md

• [ ] Hardcoded sessionId: 'default' or no sessionId on memory. All conversations share one session or sessions won't be used properly. → MEMORY.md

• [ ] Image / audio / video generation wrapped in an Agent. Binary doesn't flow through tools or the Agent's output formatter. Use the provider's native single-call node directly. → n8n-ai-nodes-official anti-patterns

• [ ] Agent + Switch to route on natural-language input when Text Classifier (@n8n/n8n-nodes-langchain.textClassifier) is one node with N built-in branches. → n8n-ai-nodes-official

• [ ] Agent (or Basic LLM Chain + structured output parser) used to pull fields out of a blob of text when Information Extractor (@n8n/n8n-nodes-langchain.informationExtractor) is one node with a typed schema, no tools, no system prompt. Use the Agent only when the extraction needs tool calls or multi-turn reasoning. → n8n-ai-nodes-official

First-class n8n Agents: quality

• [ ] Node tool without a description in its JSON. It gets a generic "Execute node" text and is picked badly. → TOOLS.md

• [ ] Tool expression input that reads trigger or upstream data ($('Webhook'), $json.<event field>). Tools only see the model's arguments, so it's empty or wrong. → TOOLS.md "What expressions can see"

• [ ] Vague useWhen on sub-agents or vector stores, or skill descriptions that don't say "Use when ...". These are the triggers; vague ones never fire. → SKILLS.md

• [ ] Task with no timezone, or an objective that doesn't say where results go. Fires at the wrong hour, or produces output nobody sees. → TASKS.md

• [ ] Slack approval where the approver needs the values (amount, recipient). Channel cards don't show parameters. → APPROVALS.md

• [ ] MCP server without toolFilter, exposing a whole catalog when the Agent needs a few tools. → TOOLS.md

System prompts (AI Agents)

The system prompt is the load-bearing config of an agent. Severity ranges by how badly the issue degrades behavior.

• [ ] Hardcoded date / missing date in the system prompt. Stale immediately. Agents making time-sensitive decisions (deadlines, eligibility windows, "is X expired?") get wrong answers. Use Current date: {{ $now }}. → SYSTEM_PROMPT.md "Always include the current date"

• [ ] Long system prompt with per-tool usage instructions buried inside. Modular split violation. Brittle to edit, and tools become unreusable across agents. Move tool-specific instructions into each tool's description. → SYSTEM_PROMPT.md "The modular split" and n8n-ai-nodes-official "What goes in the system prompt vs the tool description"

• [ ] "You are a helpful assistant" preamble with no role specifics. Generic responses, agent has no identity. Replace with a specific role and scope. → SYSTEM_PROMPT.md "What it's for"

• [ ] No refusal/safety boundary when one is needed (the agent has tools that touch user-visible state, sends, payments). Define explicit boundaries: "only answer questions about X, redirect otherwise". → SYSTEM_PROMPT.md

Data Tables

System columns and identifiers

• [ ] Set node before Data Table Insert/Update. Map directly in per-column slots. → n8n-data-tables-official strong defaults

• [ ] id, createdAt, or updatedAt declared in create_data_table. System-managed; declaring them errors or shadows. → n8n-data-tables-official non-negotiable

• [ ] Auto-id used as a cross-system identifier. Resets on table recreate / instance migration. Add a domain ID column (arxivId, customerId, requestId). → SCHEMA_DESIGN.md "system-managed columns"

_object postfix discipline (the most error-prone area)

• [ ] Column holds stringified JSON (array, object) without the _object postfix. Readers have no contract telling them to parse. → SCHEMA_DESIGN.md "the _object convention"

• [ ] Column has _object postfix but is string type holding native (non-stringified) values. Contract violation: postfix promises stringified content. → SCHEMA_DESIGN.md

• [ ] Insert/Update writes to an _object column without JSON.stringify(...). A [object Object] literal lands in the row. → SCHEMA_DESIGN.md

• [ ] Read from an _object column without JSON.parse(...). Downstream gets a string where it expects array/object; templates and tools choke. → SCHEMA_DESIGN.md

• [ ] Sub-workflow returns _object columns as strings to callers. Storage format leaking through the interface. Parse before returning so callers receive arrays as arrays and objects as objects. → SCHEMA_DESIGN.md "Storage format ≠ interface format"

• [ ] Sub-workflow's "fresh" path stringifies to "match" the cached path. Wrong instinct. Parse the cached path so both branches return natural shapes. → SCHEMA_DESIGN.md

• [ ] _object column holds data that needs to be queried (filter on topics content, find rows by tag). Strings can't be queried structurally. Refactor to a relational child table or move to a real DB. → SCHEMA_DESIGN.md "When NOT to use _object"

Schema design

• [ ] String column acting as boolean ('yes'/'no', 'true'/'false'). Use the boolean type. → SCHEMA_DESIGN.md "Picking column types"

• [ ] Mixed casing within one table or query (createdAt AND arxiv_id together). Match camelCase to the system columns. → n8n-data-tables-official strong defaults

• [ ] Table named with snake_case, lowercase, or in singular for a set. Title Case with spaces (Papers, Recent Events). Plural for sets, singular for one-row-per-global-thing only. → SCHEMA_DESIGN.md "Naming"

• [ ] Boolean column not named affirmatively (completed instead of isCompleted). Affirmative names read better in filters and IFs. → SCHEMA_DESIGN.md

• [ ] Table with more than ~17 columns. Consider splitting; the table is probably trying to be multiple things. → SCHEMA_DESIGN.md "A healthy table"

• [ ] User-declared date column duplicates createdAt / updatedAt. Use the system columns where they fit; only add explicit date columns when the timestamp's meaning differs from "row created" or "row updated". → SCHEMA_DESIGN.md

Relational design (when the data has shape)

• [ ] Genuine parent-child data crammed into one wide table (papers + summaries in one Papers table, customers + orders in one Customers table). Split into parent + child tables, reference parent by id (paperId, customerId). → SCHEMA_DESIGN.md "Designing relationally" and n8n-data-tables-official "Relational design"

• [ ] Child arrays stored in _object columns when they need to be queried or filtered individually. Refactor to a relational child table. → SCHEMA_DESIGN.md "When NOT to use _object"

• [ ] Cascade strategy unclear (no defined behavior on parent delete). Pick one per relationship: cascade-delete via a separate Delete on children, soft-delete (archived flag), or orphan. Mixed strategies cause silent bugs. → SCHEMA_DESIGN.md "Enforce integrity in the workflow"

• [ ] Multi-table write sequence (parent Insert + child Insert) without partial-failure handling. No transactions exist; a child failure leaves the parent orphaned. Pick: compensating writes, idempotent retry with upsert + stable domain IDs, or soft state marker (status: 'pending' → 'complete'). → SCHEMA_DESIGN.md "No transactions, plan for partial failure"

• [ ] Child rows pointing at deleted parents (orphan detection missing). Either soft-delete by default, or run a periodic cleanup workflow. → SCHEMA_DESIGN.md "Watch for child without parent"

• [ ] 3+ joined tables with transactional writes. Past Data Tables' wheelhouse. Use a real SQL DB. → SCHEMA_DESIGN.md "When Data Tables are the wrong tool"

Operation gotchas

• [ ] Multi-column filter without explicit matchType: 'allConditions'. Defaults to anyCondition (OR) in some versions; surprising when intent is AND. → OPERATIONS.md "matchType"

• [ ] Get without alwaysOutputData: true followed by an IF that branches on result presence. No-match produces zero items, the IF doesn't fire at all. → DEDUP_PATTERNS.md "Pattern 3: Get + IF"

• [ ] Get + IF branches produce different JSON shapes (cached path vs. freshly-processed path). Downstream $json.x resolves to the wrong field depending on which fired. Insert a Set/NoOp anchor to normalize. → n8n-expressions-official "Combine Inputs convergence"

• [ ] Update with no match silently does nothing. No error. Either follow with a Get to confirm, or use upsert for create-or-update. → OPERATIONS.md "Update with no match"

• [ ] Insert with no dedup in a workflow that can re-fire (webhook retry, scheduled re-run). Creates duplicates. Use upsert, rowNotExists + Insert, or upstream dedup. → DEDUP_PATTERNS.md

• [ ] returnAll: false without explicit limit. Defaults to 50 in many versions; downstream may expect more and silently truncate. → OPERATIONS.md "returnAll"

• [ ] Plain "have I seen this value?" dedup using a Data Table. The Remove Duplicates node ("items seen in previous executions" mode) handles this with no schema. Reach for Data Tables only when the dedup state needs to be queryable, has row-level logic, or has TTL/per-tenant scoping. → DEDUP_PATTERNS.md

• [ ] Idempotency-key workflow without TTL cleanup. Markers shouldn't live forever; run a daily cleanup. → DEDUP_PATTERNS.md "Idempotency keys"

Mapping mode and evolution

• [ ] autoMapInputData mode without a stable 1:1 between upstream field names and column names. Drift on either side silently breaks the mapping. Default to defineBelow. → SCHEMA_DESIGN.md "Mapping mode"

• [ ] Workflow attempts an in-place column type change. Not supported. Add a new column, copy via workflow, drop the old, rename the new. → SCHEMA_DESIGN.md "Schema evolution"

Wrong tool for the job

• [ ] Cross-app shared data in Data Tables. Awkward to query from outside n8n. Use a real DB. → SCHEMA_DESIGN.md "When Data Tables are the wrong tool"

• [ ] Millions of rows or write-heavy volume in a Data Table. Performance degrades; use a real DB. → SCHEMA_DESIGN.md

Binary handling

• [ ] Reading file content from $json instead of $binary. → BINARY_BASICS.md

• [ ] Agent tool returning raw binary directly. Tool output is JSON-only. Upload to storage, return key/URL in JSON. → AGENT_TOOL_BINARY.md

• [ ] Uploaded chat files passed to a tool via fromAi. fromAi doesn't carry binary. Pre-stage to storage, inject keys in the system prompt. → AGENT_TOOL_BINARY.md

• [ ] Binary lost after a JSON transform. Use Merge to combine the JSON output with the binary stream. → MERGE_FOR_CONTEXT.md

• [ ] Image sent to a chat surface from raw $binary. Chat surfaces need a URL-referenced image (or platform-native file upload). → CDN_REQUIREMENT.md

Public trigger auth

• [ ] Webhook trigger with parameters.authentication: 'none' even on read-only / lookup paths. The URL is publicly callable; no auth invites abuse (rate-limit exhaustion, scanning, scraping). Use Basic / Header / JWT auth unless the workflow is genuinely meant to be public. (If the action is sensitive, this is a MUST FIX, see above.) → n8n-credentials-and-security-official

• [ ] Form Trigger fronting a sensitive action with no auth gate. Forms that create accounts, mutate data, send external messages, or trigger paid operations should require real auth at the trigger. An obscure URL is not auth. → n8n-credentials-and-security-official

Webhook / Respond to Webhook

• [ ] Fallible nodes with no error path. HTTP / DB / API / file nodes need output(1) wired to a 5xx Respond. Without it, the failure is unhandled: the workflow halts and the caller gets whatever n8n's generic error response is (no controlled status code, no useful body), and the operator only learns about it if a workflow-level error workflow is configured separately. → n8n-error-handling-official non-negotiables and API_WORKFLOWS.md

• [ ] Error response returns 200. Caller's HTTP client treats it as success; downstream error handling never fires. Always 4xx (caller error) or 5xx (server error). → RESPONSE_SHAPES.md

• [ ] Generic 500 for every failure. Validation errors should be 400, auth issues 401/403, conflicts 409, rate limits 429. Caller can't distinguish their bug from your outage. → RESPONSE_SHAPES.md

• [ ] respondWith: 'json' body using JSON.stringify(...) instead of an object literal (produces double-encoded body). → WEBHOOK_NODES.md

HTTP Request

• [ ] Headers set both via headerParameters and a credential's httpHeaderAuth. They conflict. → HTTP_NODES.md

• [ ] Untrusted external API call without options.timeout set. Runtime default is 5 minutes; a hung request stalls the workflow. → HTTP_NODES.md

Schedule trigger

• [ ] Business-critical schedule with no explicit timezone at the workflow level. DST and instance moves cause timing shifts. → TRIGGER_NODES.md

• [ ] Schedule-triggered workflow not idempotent for missed-run scenarios. Restarts or downtime can miss runs. → TRIGGER_NODES.md

NICE TO HAVE

Naming

• [ ] Workflow name doesn't follow verb-first pattern (Send weekly customer report vs. Customer report sender). → NAMING_CONVENTIONS.md

• [ ] Untagged sub-workflow (missing subworkflow, a domain tag, or tool). Tags are the discovery mechanism; an untagged sub-workflow won't surface under any tags filter. → NAMING_AND_DISCOVERY.md

Readability

• [ ] Canvas over 7 top-level boxes (n8n 2.28+) whose logical steps aren't grouped into node groups. Boxes are the trigger, each group, and each ungrouped node; sub-nodes and stickies don't count. Collapsed groups make the canvas read as steps, not nodes. Skip on older instances, where node groups don't exist. → NODE_GROUPS.md

• [ ] Group with no description, or one written to restate its title. Groups are created collapsed, so title plus description is all a reader gets. Max 145 characters. → NODE_GROUPS.md

• [ ] An Agent grouped apart from its sub-nodes. Any non-main connection crossing the boundary drops the group; all in or all out. → NODE_GROUPS.md

• [ ] Sticky note used to fake a group instead of a node group. Stickies annotate (callouts, TODOs, warnings); they don't group. → SKILL.md "Readability"

• [ ] Sticky title re-states what's visible (Set, If, Set). Title with the callout's point. → SKILL.md "Readability"

• [ ] Sticky color used for variety instead of meaning. Color should say what happens there (green creates or involves a person, red destructive or irreversible, blue routing, gray documentation), one meaning per color, matching whatever nearby workflows already use. → STICKY_NOTES.md

• [ ] Sticky sitting away from, or on top of, what it describes. On create, anchor it with sticky(content, [nodes]) instead of positioning by hand. A note at [0, 0] is the usual tell, though saved JSON can't distinguish an omitted position from a deliberate one, so check where it should sit rather than how it got there. → STICKY_NOTES.md

• [ ] Workflow description is empty, one sentence, or doesn't capture the why. Two sentences: what it does, why it exists. → SKILL.md "Readability"

• [ ] Code node without a one-line note explaining its purpose. → SKILL.md "Readability"

• [ ] Workflow ignores existing house style (sticky palette, naming convention used in nearby workflows). Match what's there. → SKILL.md "Readability"

Sub-workflow conventions

• [ ] Sub-workflow without a final Return Set / Edit Fields node shaping the return contract. The legitimate exception to the Set-node antipattern. → n8n-subworkflows-official "Other conventions"

• [ ] Inputs / outputs not documented in the sub-workflow's description. Field names, types, purpose. → n8n-subworkflows-official "Sub-workflow inputs and outputs"

Conventions

• [ ] Per-execution context (user identity, files, current task) buried in a static system prompt instead of injected via expressions or piecing. Hard to update. → SYSTEM_PROMPT.md "Storing the prompt"

• [ ] Generic safety boilerplate language in the system prompt for risks the model already handles. Reinforcement adds tokens without changing behavior. Reserve safety language for specific, named risks. → SYSTEM_PROMPT.md anti-patterns

Reporting findings

When reporting, group by severity, then within severity by domain. For each finding include:

• The specific node(s) or section affected.

• A one-sentence description of the issue.

• The link to the canonical skill ref for the fix.

MUST FIX
  Security
  - Node `Send Webhook`: bearer token in headerParameters value field. → n8n-credentials-and-security-official
  - Node `Lookup user`: SQL string concat with $json.email. → DATABASE_NODES.md

  Connections
  - Node `Validate input`: `connections.Validate input.main[0]` has only 1 entry, but the workflow's logic expects fan-out to 3 downstreams. → check the SDK code (likely a missed `.add(...)` call)

SHOULD FIX
  ...

A review agent should not auto-fix MUST FIX items without user confirmation: security and connection fixes have blast radius and the user should know what's changing.