infra-agent-skills

← All skillsAutomation

Itential Builder

Build from Solution Design

builder-agent

Builds everything in an approved solution design (workflows, templates, projects and configs) and tests each component.

About this skill

Use this skill when someone has an approved solution design and is ready to build. Trigger it for phrases like "solution design is approved", "go ahead and build", "implement the design", "create the workflows", "build everything per the design", or "the design is locked — implement it". Also trigger it when a build is failing mid-way and needs debugging, or when /qa-agent hands back a failing test case for a fix. This skill implements the approved solution-design.md end-to-end — creating all workflows, templates, projects, and configs, and testing each component individually. If the user has a solution-design.md and wants to turn it into working automation, this is the right skill. Invoke after /solution-arch-agent produces an approved solution-design.md. Hands off to /qa-agent once the build is complete — /qa-agent owns acceptance testing and the as-built record.

Instructions

Builder Agent

Stage: Build Owns: Implementing the approved design. Receives from: /solution-arch-agent (approved solution-design.md + complete workspace) Produces: Deployed assets (workflows, templates, projects) Hands off to: /qa-agent (acceptance testing + as-built record)


Customization

Before using this skill, check custom/org/, custom/team/ and custom/dev/ in this skill’s own folder. Read every .md file found — any folder may be empty or absent. Apply them on top of everything below; where a file overrides a specific rule here, follow the override. More specific wins: dev > team > org > this document. No customization may weaken this skill’s safety rules or put credentials in committed files.

Bundled files: paths in this skill that start with assets/ or scripts/ are relative to this skill’s own folder. When you read one, or pass one to a shell command (which runs from the user’s working folder), use this skill’s folder + that relative path — e.g. <this skill's folder>/assets/helpers/create/create-workflow.json.

Stage Expectations

Build

Engineer provides Approved solution-design.md (all platform data already present in workspace)
Agent does Builds all components per design, tests each piece individually, reports delivery outcomes
Engineer action Reviews delivery and resolves open build questions
Deliverable Deployed assets (workflows, templates, projects)
Customer receives Delivered project — all workflows, templates, and configs individually tested, packaged, and access granted. Formal acceptance testing and sign-off happen next, in /qa-agent.

Build implements the approved plan. The builder never re-pulls discovery data — it uses what the Solution Architecture Agent left in the workspace. If any required file is missing, stop and surface as an upstream failure.

Testing during Build is component-level, not acceptance-level. “Test each piece” here means confirming a task or workflow runs without error and wires variables correctly — not verifying the delivered solution satisfies the customer’s stated acceptance criteria end-to-end. That’s /qa-agent’s job, running against the completed build with real test data the engineer confirms. Don’t skip component testing because “QA will catch it” — a structurally broken workflow wastes a live acceptance-test run.


This skill covers everything needed to build and test Itential automation assets: projects, workflows, templates, and command templates.

FlowAI agents: load /flowagent before building agents. When the design has a parent agent that calls child agents, build each child as its own agent, then attach it to the parent as an agent:<child-id> tool, following “Agents as tools” in /flowagent. Don’t wrap the agents in a workflow or runCode unless the design says to. The workflow and childJob steps below apply to workflow components only. When you test, run the child on its own and then through the parent. Record both session IDs, and check the child’s answer, not just the parent’s tool status.

Phases built agent first (Section C of the design says “Agent first”): build the agent and run the test sessions the implementation plan names. Once they meet the plan’s definition of proven and the engineer agrees, record the session IDs in use-case-memory.md and hand off to /flowagent-to-spec for that phase. Its <phase>-converted-spec.md comes back through /solution-arch-agent design-only before the deterministic assets are built — don’t build them from the agent directly. Meanwhile, build the delivery’s other phases as designed.

Workspace Contract

The builder receives a complete workspace. All discovery data is already present. Solution-design (or setup for explore mode) has already pulled everything.

Required files (must exist before build starts):

{use-case}/
  .auth.json              ← auth token
  .env                    ← credentials (for re-auth if token expires)
  openapi.json            ← API reference (pulled by solution-arch-agent or explore)
  tasks.json              ← task catalog (pulled by solution-arch-agent or explore)
  apps.json               ← app/adapter type names (pulled by solution-arch-agent or explore)
  adapters.json           ← adapter instances (pulled by solution-arch-agent or explore)
  applications.json       ← app health (pulled by solution-arch-agent or explore)

May also exist (spec-contingent):

  use-case-memory.md      ← living context: IDs, decisions, gotchas, open items — READ THIS FIRST
  customer-spec.md        ← approved HLD (Requirements)
  feasibility.md          ← approved feasibility assessment
  customer-context.md     ← business rules (if provided)
  solution-design.md      ← approved Solution Design / LLD
  devices.json            ← device inventory
  workflows.json          ← existing workflows
  device-groups.json      ← device groups
  task-schemas.json       ← fetched on demand during build (append-only, never pre-populated)
  test-report.md          ← if returning from /qa-agent with a failing test case — has the exact case ID, expected vs actual, and evidence

The builder NEVER re-pulls bootstrap or discovery data. If tasks.json, apps.json, or adapters.json is missing, stop and tell the user — that’s an upstream failure, not something to silently fix.

Exception — .auth.json bootstrap: If .auth.json is missing but .env exists with AUTH_METHOD=oauth, CLIENT_ID, and CLIENT_SECRET, the builder MUST authenticate and create .auth.json before proceeding — do NOT stop and report an upstream failure. See the Bootstrap Authentication section below.

The only API calls the builder makes are:

  • Auth bootstrap — POST /oauth/token when .auth.json is missing (see below)
  • Create — POST workflows, templates, projects
  • Update — PUT to edit assets
  • Test — POST jobs/start, GET job status
  • Schema fetch — task schemas not yet in task-schemas.json (append to file after fetching)
  • Re-auth — if token expires, use .env to refresh .auth.json
  • FlowAI agents — create and update agents, read tools (GET /tools/...), start test sessions, and read sessions and their messages. Never complete a WorkCenter work item; that’s the assigned person’s job.

Bootstrap Authentication

When .auth.json is missing but .env has AUTH_METHOD=oauth with CLIENT_ID and CLIENT_SECRET, authenticate automatically before proceeding.

The correct Itential SaaS/Cloud OAuth endpoint is:

POST {PLATFORM_URL}/oauth/token
Content-Type: application/x-www-form-urlencoded

Body (form-encoded, NOT JSON — JSON returns 415):

grant_type=client_credentials&client_id={CLIENT_ID}&client_secret={CLIENT_SECRET}

Critical:

  • Content-Type MUST be application/x-www-form-urlencoded — NOT application/json. Sending JSON returns HTTP 415.
  • The /login endpoint does NOT support OAuth client credentials on SaaS instances — always use /oauth/token.
  • On success, write .auth.json with the token so all subsequent API calls just work.

Build Lifecycle

0. Memory file              → create or read {use-case}/use-case-memory.md
1. Decompose                → identify parent/child split before writing any code
2. Create project           → container for all assets
3. Discover tasks           → search tasks.json, fetch schemas
4. Build children first     → each child workflow independently testable
5. Build templates          → Jinja2 (config gen) or TextFSM (output parsing)
6. Build command templates  → MOP pre/post checks with validation rules
7. Build orchestrator last  → parent wires tested children via childJob
8. Add assets to project    → move/copy into the project
9. Set project membership   → resolve spec members, PATCH immediately after import
10. Test each component     → jobs/start, check results (component-level, not acceptance-level)
11. Debug                   → check job.error, filesystem-first
12. Reconcile               → diff built vs designed, update artifacts
13. Update memory file      → record IDs, decisions, gotchas, test results, open items
14. Update this skill       → if you hit a platform behavior not documented here, add it before closing out
15. Hand off to /qa-agent   → build is complete; real IDs are in solution-design.md §D

Step 0 — memory file:

At the start of every session, check for {use-case}/use-case-memory.md ({use-case} = use-cases/<use-case-name>/ in the working folder):

  • Exists → read it before doing anything else. It tells you the platform, project ID, what’s already built, decisions made, and open items. Don’t re-discover what’s already documented.
  • Missing → create it now from assets/helpers/use-case-memory.md template. Fill in Platform URL, Stage: build, Status: active immediately.

Step 13 — update memory file after every session:

Before closing out any build session, update use-case-memory.md with:

  • Any new asset IDs (project ID, workflow UUIDs, transformation IDs, adapter names)
  • Any architectural decisions made and why
  • Any gotchas hit and how they were fixed
  • Test results (date, what was tested, outcome)
  • Updated open items list
  • Stage and Status if they changed — mid-build, Stage stays build; only update it to test at the Step 15 handoff

The memory file is what makes it possible to pick up a use-case after weeks without re-discovering everything from scratch.

Step 14 — how to update this skill:

  • New platform behavior (error shape, field constraint, task gotcha) → add detail to the relevant body section (### query, ### childJob, ### Projects, etc.), then add a one-liner to the Gotchas pre-flight list under the right category.
  • New pattern or workflow recipe → add to ## Workflow Patterns and, if the pattern is reusable, export the project from the platform and, in a clone of the builder-skills repo, save it to the shared library at assets/helpers/assets/ (CI bundles it into this skill as assets/helpers/assets/). Add a row to the Helper Templates table in this file pointing to it.
  • Do NOT create a new top-level section for a single finding — put it where a builder would look when working on that topic.

Step 15 — hand off to /qa-agent: Once every component has been individually tested and solution-design.md Section D has real IDs (project ID, workflow IDs) instead of placeholders, the build is complete. Update use-case-memory.md to Stage: test before ending the session. Tell the engineer the build is done and route to /qa-agent for acceptance testing and the as-built record — don’t write as-built.md here.

If /qa-agent hands back a failing test case: fix the specific issue it identifies (it gives you the case ID, expected vs. actual, and evidence — a job ID or static-check output). Don’t re-examine the whole build; the failure report tells you exactly what broke. Once fixed, tell the engineer so /qa-agent can re-run just that case.


Pending Integration Pattern

Use this pattern when the solution design includes ⚠ Stub integrations — adapters that are required but not yet installed on the platform, or whose connection details (hostname, auth method, credentials) are still being confirmed by the customer.

Goal: deliver a fully buildable, testable project now. The customer sees real workflows and real progress. When the adapter arrives, swapping in the real task is a single targeted update.

Stub Workflow (stub-{integration-name})

Build one stub workflow per pending integration. It exercises one core connectivity action — enough to prove the integration wires up correctly end-to-end when the adapter is provisioned.

Structure:

workflow_start
  → buildPayload    (newVariable — assembles minimum input for the core action)
  → callIntegration (newVariable placeholder — sets {integrationName}Status = "pending_adapter")
       ↓ error
  → workflow_end
workflow_end

Rules:

  • buildPayload — newVariable task. Assembles the minimum required input per the integration-model-{name}.json (e.g., for Slack: {channel, text}). No adapter dependency — always runnable without a real adapter.
  • callIntegration — newVariable placeholder. Sets {integrationName}Status = "pending_adapter". Use the hex task ID that the real adapter task will occupy. Error transition pre-wired to workflow_end.
  • One input variable: dryRun (boolean, default true). Stub ignores it; activated workflow can use it to skip side effects during testing.
  • Add all stub workflows to the same project as the main delivery workflows.

Choosing the core action:

  • Prefer the simplest write/action that validates auth end-to-end (post a message, launch a job, write a secret)
  • If the integration is read-only in this use case, use a lightweight read (get current user, health check)
  • Derive the request payload from integration-model-{name}.json — that file is the contract

Placeholder Task (in any workflow)

When a workflow needs to call an adapter that isn’t installed, replace the real adapter task with a newVariable placeholder — same position, same task ID, same transitions:

  1. Use a newVariable task in the exact slot the real adapter task will occupy
  2. Assign the same hex task ID the real task will use when activated
  3. Set variable: {integrationName}Status = "pending_adapter" — machine-readable pending state
  4. Wire all transitions identically (including the error transition) to how the real task will be wired

As-Built: Activate Integration Section

For every pending integration, include a dedicated section in as-built.md:

## Activate: {Integration Name}

When the {AdapterType} adapter is provisioned:

1. Confirm adapter instance name:
   jq '.results[] | select(.package_id | test("{name}";"i")) | {id,state}' adapters.json

2. Replace task `{taskId}` in workflow `{workflowId}` with:
   (complete replacement task JSON — all fields pre-filled from integration-model-{name}.json,
    app and locationType set to the adapter type name, incoming variables wired from the workflow)
   Only field left blank: `adapter_id`. Fill from step 1.

3. Verify exact field names from the live schema before activating:
   POST /automation-studio/multipleTaskDetails?dereferenceSchemas=true
   body: {"tasks": [{"name": "{taskName}", "app": "{appType}"}]}

This pattern works for both standard Itential adapters (EmailOpensource, Slack) and custom OpenAPI virtual integrations. The integration-model-{name}.json file is the contract in both cases.


Guides

Building a Gateway4 → Gateway5 migration? Follow assets/helpers/gateway-migration/conversion-guide.md for each item’s pattern (inventory with broker actions, services imported through Gateway Manager, task-by-task rewire, re-pointing every consumer of a Gateway4 output), and import the migrated workflows as a new project — never edit the originals.

Guide 1: Build a workflow end-to-end

Follow these steps in order. Do not skip any step.


Before writing task JSON, extract a real example from assets/helpers/assets/ — do not invent task structure from memory. (assets/helpers/create/ is for API wrappers only — project/workflow creation endpoints, not task bodies.)

# 1. Find which asset project matches your use case
ls assets/helpers/assets/

# 2. Extract the workflow most similar to what you're building
jq '[.components[] | select(.type=="workflow")] | .[].document.name' \
  assets/helpers/assets/vendor-servicenow.json

# 3. Read its full task map — this is your reference
jq '[.components[] | select(.type=="workflow") | select(.document.name | test("WORKFLOW_NAME"; "i"))] | first | .document | {tasks, transitions}' \
  assets/helpers/assets/vendor-servicenow.json

# 4. Extract the specific task type you need
jq '[.components[].document.tasks // {} | to_entries[] | select(.value.name == "TASK_NAME")] | first | .value' \
  assets/helpers/assets/vendor-servicenow.json

Replace assets/helpers/assets/vendor-servicenow.json with whichever asset file best matches your use case:

  • Adapter tasks (ServiceNow, Infoblox) → assets/helpers/assets/vendor-servicenow.json, assets/helpers/assets/vendor-infoblox-nios-ddi.json
  • Network device tasks (CLI, MOP) → assets/helpers/assets/vendor-cisco-ios.json, assets/helpers/assets/vendor-arista-eos.json, assets/helpers/assets/vendor-juniper-junos.json
  • IPAM/inventory → assets/helpers/assets/vendor-netbox.json
  • Data transformations → assets/helpers/assets/itential-platform-data-manipulation.json
  • Config management (RunCommandTemplate, itential_cli) → assets/helpers/assets/itential-platform-configuration-management.json
  • LCM action workflows → assets/helpers/assets/lcm/lcm-vxlan-fabric-services-project.json

Step 0: Decompose before you build.

Before writing any JSON, identify the parent/child split from the solution design. Ask for each phase:

  • Can this phase be run and tested on its own? → Child workflow
  • Does it loop over multiple items (devices, records)? → Child workflow with loopType
  • Is it reusable across other use cases? → Child workflow
  • Is it a simple sequential step with no independent test value? → Task in orchestrator

Build order is always: children first, orchestrator last. The orchestrator is just childJob calls to tested children — it should not contain raw adapter tasks unless there is no logical way to split.

Step 0a: Before building a per-item transform loop, default to runCode + Enable Query instead of chaining WorkFlowEngine utility tasks.

If a phase does ANY of the following, don’t reach for forEach + query×N + evaluation + newVariable + merge + push/objectToString/join/makeData — build it as a single runCode (Python, GatewayManager) task instead:

  • Loops over a list, extracting 2+ fields per item
  • Branches per item on a field value (manufacturer, status, type) to pick between two or more output shapes
  • Reshapes each item into a different structure before handing it to the next task

When NOT to use runCode: the per-item work needs to call another platform task (adapter, app method, childJob) for each item — runCode only runs Python against the data payload it’s given; it cannot invoke other Itential tasks. Use forEach+childJob/adapter-task for that instead.

See the ### runCode section under Utility Tasks below for the collapse story, field reference, and the exact jq commands to extract a verified real example before building any transform loop.

Read a full workflow from asset projects before building any multi-workflow solution:

# Parent → childJob → evaluation pattern
jq '[.components[] | select(.type=="workflow") | select(.document.name | test("Upgrade|Runner"))] | first | .document | {name,tasks,transitions}' \
  assets/helpers/assets/vendor-cisco-ios.json

# childJob loop with data_array
jq '[.components[] | select(.type=="workflow") | select(.document.name | test("Chunk|Loop"))] | first | .document | {name,tasks,transitions}' \
  assets/helpers/assets/itential-platform-configuration-management.json

Step 1: Find tasks. Search tasks.json for the tasks you need:

jq '.[] | select(.name | test("keyword"; "i")) | {name, app, type, location, canvasName, displayName}' {use-case}/tasks.json

Step 2: Resolve adapter app names. For adapter tasks, the app in tasks.json is WRONG. Look up the correct name:

jq '.[] | select(.name | test("keyword"; "i")) | {name, type}' {use-case}/apps.json

Also get the adapter instance name:

jq '.results[] | select(.package_id | test("keyword"; "i")) | {id, state}' {use-case}/adapters.json

You now have three values: app (from apps.json), adapter_id (from adapters.json .id), and displayName (from tasks.json).

Step 3: Fetch task schemas. Get the full input/output schema for every task you’ll use:

POST /automation-studio/multipleTaskDetails?dereferenceSchemas=true
{
  "inputsArray": [
    {"location": "Adapter", "pckg": "Servicenow", "method": "createChangeRequest"},
    {"location": "Application", "pckg": "WorkFlowEngine", "method": "query"}
  ]
}

Use the pckg value from apps.json (Step 2), NOT tasks.json. Save the response to {use-case}/task-schemas.json.

Step 4: Map schema to workflow task JSON. For each task, transform the schema into a workflow task:

Schema response:

{
  "name": "createChangeRequest",
  "variables": {
    "incoming": {
      "body": {"type": "object", "description": "Request body"}
    },
    "outgoing": {
      "result": {"type": "object", "description": "Response"}
    }
  }
}

Becomes this workflow task (extract a real adapter task from an asset project first — e.g. jq '[.components[].document.tasks // {} | to_entries[] | select(.value.location == "Adapter")] | first | .value' assets/helpers/assets/vendor-servicenow.json):

{
  "a1b2": {
    "name": "createChangeRequest",
    "canvasName": "createChangeRequest",
    "summary": "Create Change Ticket",
    "description": "Creates a ServiceNow change request",
    "location": "Adapter",
    "locationType": "Servicenow",
    "app": "Servicenow",
    "type": "automatic",
    "displayName": "ServiceNow",
    "variables": {
      "incoming": {
        "body": "$var.e1a1.merged_object",
        "adapter_id": "$var.job.adapter_id"
      },
      "outgoing": {
        "result": null
      },
      "error": "",
      "decorators": []
    },
    "groups": [],
    "actor": "Pronghorn",
    "scheduled": false,
    "nodeLocation": {"x": 700, "y": 600}
  }
}

Mapping rules:

  • task ID (the key in tasks) → hex only, [0-9a-f]{1,4} (e.g. d4e5) — never letters g–z, and use the same ID in transitions
  • name, canvasName → from tasks.json
  • app, locationType → from apps.json (NOT tasks.json)
  • displayName → from tasks.json
  • location → "Adapter" or "Application" (from tasks.json)
  • type → from tasks.json directly — do not guess. It is per-task, not per-app. Read it alongside name, app, location, and canvasName: jq '.[] | select(.name == "taskName") | {name, app, type, canvasName, location}' tasks.json
  • actor → "Pronghorn" for all tasks except childJob (which uses "job")
  • incoming → each schema key becomes a variable. Wire with $var for top-level values
  • outgoing → set to null (capture later with $var.taskId.outVar)
  • Add adapter_id to incoming for adapter tasks (not in schema, always required)
  • Add error and decorators to variables block

Step 5: Handle object inputs. If a task’s incoming variable is type: "object" (like body), you CANNOT put $var references inside it — they won’t resolve. Use a merge task before it:

{
  "e1a1": {
    "name": "merge",
    "canvasName": "merge",
    "summary": "Build Request Body",
    "app": "WorkFlowEngine",
    "type": "operation",
    "variables": {
      "incoming": {
        "data_to_merge": [
          {"key": "short_description", "value": {"task": "job", "variable": "short_description"}},
          {"key": "description", "value": {"task": "job", "variable": "description"}}
        ]
      },
      "outgoing": {"merged_object": null}
    },
    "actor": "Pronghorn"
  }
}

Then wire the adapter task’s body to "$var.e1a1.merged_object".

Step 6: Handle opaque schemas. Some task schemas show body: {type: "object"} with no inner field details. The adapter validates internally. To discover required fields:

  1. Try creating with minimal fields — the error message lists what’s missing (e.g., "must have required property 'summary'")
  2. Check openapi.json for the adapter’s endpoint schema
  3. Call the adapter directly: POST /{adapter_id}/{method} with {} body — read the validation error

Step 7: Wire transitions. Every adapter task needs BOTH success and error transitions:

"transitions": {
  "a1b2": {
    "b2c3": {"type": "standard", "state": "success"},
    "ef01": {"type": "standard", "state": "error"}
  }
}

If both success and error need to reach workflow_end, route error to an intermediate newVariable task first (JSON can’t have duplicate keys).

Step 8: Add inputSchema/outputSchema. List all job variables the workflow expects as input and produces as output.

Step 9: Pre-submit checklist.

  • Every task has a non-empty description field — schema-required on every task, easy to forget, and POST /workflow_engine/workflows/validate will reject a document missing it
  • Task IDs are hex-only ([0-9a-f]{1,4}) — every key in tasks and every key and target in transitions; only workflow_start and workflow_end are exempt. A letter g–z (e.g. g1g1) makes the import fail. Verify with python3 -c "import json,re; d=json.load(open('workflow.json')); bad=[k for k in d['tasks'] if k not in ('workflow_start','workflow_end') and not re.match(r'^[0-9a-f]{1,4}$', k)]; print(bad or 'OK')" rather than eyeballing the task list. A non-hex ID (e.g. g007, using a letter outside a–f) doesn’t error at creation time — it causes intermittent, hard-to-diagnose $var resolution failures later (evaluation operands silently drilling into the wrong value, string references passing through as unresolved literal text) that look like unrelated wiring bugs.
  • Every workflow document has "canvasVersion": 3 at the top level. Copy the value from assets/helpers/create/create-workflow.json; do not guess. This applies to new workflows, updates and imports. Without it, POST /workflow_engine/workflows/validate returns isValid: false (“Task has nodeLocation but canvasVersion is not 3”), and an import still succeeds but stores the workflow as canvas version 1 (the old canvas).
  • app and locationType values come from apps.json .name, NOT tasks.json and NOT the adapter instance name (e.g., EmailOpensource not email)
  • adapter_id is the adapter instance name (e.g., email), NOT the type name
  • adapter_id values come from adapters.json .results[].id — NEVER from the spec’s adapter identity table. The spec is a design document; adapters.json is the source of truth for the target environment.
  • canvasName values come from tasks.json canvasName field
  • Every adapter task has adapter_id in incoming
  • No newVariable or query task exists only to hold or extract a value, and job variables are limited to workflow inputs, outputs and shared values (see “Keep variables to a minimum”)
  • Every adapter task has an error transition — /workflow_engine/workflows/validate does not check for this; missing error transitions pass validation and hang the job at runtime
  • evaluation tasks have both success AND failure transitions — /workflow_engine/workflows/validate does not check for this either
  • evaluation operators are from the closed enum (contains, !contains, <, <=, >, >=, ==, !=) — no others exist. /workflow_engine/workflows/validate does not enforce this enum on a workflow’s embedded evaluation tasks; an invalid operator passes validation and fails silently at runtime (finish_state: failure, no error message)
  • evaluation operand_2 literal values containing regex metacharacters (., (, ), [, ], ?, +, *, |) are properly escaped, OR stored in a newVariable constant-holder task to avoid incomingRefs cache issues after API PUT
  • No $var.<taskId>.<out> references inside nested forEach bodies — use $var.job.<varName> instead
  • Incoming variable types match task schema exactly (arrays for to/cc/bcc, numbers for page/pageSize, etc.)
  • No $var references inside nested objects (use merge/makeData)
  • merge uses "variable", childJob uses "value"
  • No {task:"job", variable:"x"} in merge/childJob for workflow-internal variables — {task:"job"} refs add x to inputSchema.required, prompting operators for values that should be internal. Use the producing task ref instead (query→return_data, newVariable→value, makeData→output, merge→merged_object)
  • If a query downstream of a childJob returns null despite the child succeeding: "obj": "$var.<childJobId>.job_details" is a stale incomingRefs entry (not a platform-version issue). Fix: insert a merge task between childJob and query using {"task": "<childJobId>", "variable": "job_details"} in data_to_merge, then point obj to $var.<mergeId>.merged_object (see Guide 4)
  • childJob has actor: "job", all others have actor: "Pronghorn"
  • workflow_end transition is empty {}
  • Canvas layout follows the vertical spacing convention — non-forked sequences on a constant-x spine, fork branches offset to spine±264 and stay in their own column until convergence
  • No transition lines cross task nodes (the spine column is empty between a fork and its convergence point)
  • Sequential y-delta ~108px (tight grid)
  • LCM Create actions only: the instance-write merge task’s data_to_merge covers every field in the resource model’s schema.required array — missing even one field causes an instance write failure after provisioning (resources are orphaned from LCM). Read the model’s schema.required before building the merge task: jq '.schema.required' assets/helpers/assets/lcm/<model>.json
  • ViewData manual tasks: type MUST be "manual" — view is a top-level field; incoming.variables is present (even if {}); displayName: "Tools", no actor field (manual tasks never take one). A wrong type on a manual-view task can crash POST /workflow_engine/workflows/validate (HTTP 500) or produce a contradictory actor-required error on a task that never needs one — fix type first, don’t chase either symptom.
  • restCall downstream query: path targets body field directly (e.g., "access_token") — NOT "response.access_token" (restCall has no wrapper, unlike adapter tasks)
  • childJob loop: if child workflow has inputSchema.required fields beyond what each data_array element contains, use the forEach enrichment pattern (forEach → merge → arrayPush) to add shared fields into each element before the childJob loop; set variables: {} on the childJob
  • forEach body: incoming contains ONLY data_array (no job_id); loop body tasks have no external error transitions; last body task has an empty {} transition; $var.job.<varName> inside loop body instead of $var.<taskId>.<output>
  • makeData with childJob-sourced merge: if a merge task references a childJob variable, do NOT wire that merge’s merged_object into makeData.incoming.variables — use query to extract individual values first
  • objectToString: omit replacer/space from incoming if unused — do NOT pass replacer: [] (silently whitelists zero properties, output is always {}) or null (validate-time type warning)

Complete working example: Read the ServiceNow “Create Change Request” workflow before building — it demonstrates merge → adapter create → query → adapter update with error transitions:

jq '[.components[] | select(.type=="workflow") | select(.document.name | test("Create Change"))] | first | .document' \
  assets/helpers/assets/vendor-servicenow.json

How the example works — what each task does and why:

workflow_start → e1a1 (merge) → a1b2 (createChangeRequest) → b2c3 (query) → c3d4 (updateChangeRequest) → workflow_end
                                  ↓ error                                      ↓ error
                                ef01 (newVariable) ────────────────────────────→ workflow_end
Task ID Task Why it’s there Key fields
e1a1 merge Builds the body object. $var can’t resolve inside nested objects, so merge assembles the object from individual variables. data_to_merge uses "variable" (NOT "value"). Needs at least 2 items.
a1b2 createChangeRequest Adapter call. body wired to $var.e1a1.merged_object (merge output). app/locationType from apps.json (Servicenow), NOT tasks.json (ServiceNow). adapter_id added manually (not in schema). type: "automatic".
b2c3 query Extracts the change ID from the adapter response. query: "response.id" — adapters transform responses, don’t assume native API shape.
c3d4 updateChangeRequest Second adapter call using the extracted ID. changeId wired from $var.job.changeId (set by query’s outgoing).
ef01 newVariable Error handler. Adapter error transitions route here. Exists because JSON can’t have duplicate keys — can’t route both success and error to workflow_end from the same task.

Field mapping — where each value comes from:

Workflow task field Source Example
name tasks.json .name createChangeRequest
canvasName tasks.json .canvasName createChangeRequest (can differ: arrayPush→push)
app apps.json .name (adapter type name) Servicenow, EmailOpensource (NOT email, NOT ServiceNow from tasks.json)
locationType Same as app for adapters, null for applications Servicenow, EmailOpensource
displayName tasks.json .displayName ServiceNow, email
location tasks.json .location Adapter or Application
type tasks.json .type — read directly, do not guess (per-task, not per-app) varies
actor "Pronghorn" always, except childJob which uses "job" Pronghorn
adapter_id adapters.json .results[].id (adapter instance name) servicenow-prod, email — this goes in incoming, NOT in the task-level app field
incoming vars From task schema (multipleTaskDetails) body, changeId
outgoing vars From task schema, set to null result

Guide 2: Debug a failed job

GET /operations-manager/jobs/{jobId} → check data.status. If "error", read data.error[] (.task = failing task ID, .message.IAPerror.displayString = human-readable error). See the full symptom → cause → fix table under Debug Failed Jobs (Testing & Debugging section below) — it’s the canonical, most complete version.

Guide 2b: Work with any unfamiliar adapter task

Follow Guide 1 Steps 1-6 for discovery. Quick reference for the lookup commands:

# Step 1 — find the task
jq '.[] | select(.app | test("meraki";"i")) | {name, app, displayName}' {use-case}/tasks.json

# Step 2 — get the correct app name (tasks.json app field is often wrong for adapters)
jq '.[] | select(.name | test("meraki";"i")) | {name, type}' {use-case}/apps.json

# Step 2 — get the adapter instance name
jq '.results[] | select(.package_id | test("meraki";"i")) | {id, state}' {use-case}/adapters.json

# Step 3 — fetch the task schema
# POST /automation-studio/multipleTaskDetails?dereferenceSchemas=true
# {"inputsArray": [{"location": "Adapter", "pckg": "<app from apps.json>", "method": "<task name>"}]}

You now have three values: app (from apps.json), adapter_id (from adapters.json .id), displayName (from tasks.json). Two things to pay extra attention to beyond Guide 1:

Enforce data types from the schema. When the schema says "type": "array", you MUST pass an array — even for single values:

  • "to": "user@example.com" → WRONG. Use "to": ["user@example.com"]
  • "pageSize": "100" → WRONG if schema says number. Use "pageSize": 100
  • "cc": "" → OK only if schema allows string; if array, use "cc": []

Always check task-schemas.json for the exact type of each incoming field before wiring.

Inspect the actual response before wiring a query path. Adapter responses are transformed — they do not match the native API’s structure. After a successful test run:

  1. GET /operations-manager/jobs/{jobId} — find the task in data.tasks by its task ID
  2. Read the task’s outgoing variables — that is the real response object
  3. Use jq to explore: jq '.data.tasks["a1b2"]' job.json
  4. Wire the query path from what you see — not from the upstream API docs

End-to-end sequence:

1. tasks.json search   → found "getDevice", app "networkAdapter"
2. apps.json lookup    → correct app name is "NetworkAdapter" (capital N)
3. adapters.json       → adapter_id is "network-prod-1"
4. multipleTaskDetails → incoming: {deviceId: string}, outgoing: {result: object}
5. Build + test        → job completes
6. Inspect job         → result is {"response": {"hostname": "...", "model": "..."}}
7. Wire query path     → "response.hostname" (NOT "result.hostname" or "data.hostname")

Guide 3: Add a task to an existing workflow

Step 1: Extract the task structure from an asset project that uses the same task type:

# Adapter task (e.g., ServiceNow)
jq '[.components[].document.tasks // {} | to_entries[] | select(.value.location == "Adapter")] | first | .value' \
  assets/helpers/assets/vendor-servicenow.json

# Application task (WorkFlowEngine)
jq '[.components[].document.tasks // {} | to_entries[] | select(.value.app == "WorkFlowEngine" and .value.name != "childJob")] | first | .value' \
  assets/helpers/assets/itential-platform-configuration-management.json

# childJob
jq '[.components[].document.tasks // {} | to_entries[] | select(.value.name == "childJob")] | first | .value' \
  assets/helpers/assets/vendor-cisco-ios.json

Step 2: Fill in the fields using the mapping rules from Guide 1 Step 4.

Step 3: Generate a hex task ID (e.g., d4e5) — must be [0-9a-f]{1,4}.

Step 4: Add the task to tasks and add transitions. Remember error transitions on adapter tasks.

Step 5: Update via PUT /automation-studio/automations/{id} with {"update": {...}}.

Guide 4: Build a childJob (parent calls child workflow)

childJob has two modes. Both are tested and verified on a live platform.

Mode A: Single child — pass variables with {"task","value"}

The parent passes specific variables to one child workflow run.

Parent childJob task:

{
  "a1a1": {
    "name": "childJob",
    "canvasName": "childJob",
    "summary": "Run Single Child",
    "location": "Application",
    "locationType": null,
    "app": "WorkFlowEngine",
    "type": "operation",
    "displayName": "WorkFlowEngine",
    "variables": {
      "incoming": {
        "task": "",
        "workflow": "My Child Workflow",
        "variables": {
          "deviceName": {"task": "job", "value": "targetDevice"},
          "action": {"task": "static", "value": "validate"}
        },
        "data_array": "",
        "transformation": "",
        "loopType": ""
      },
      "outgoing": {"job_details": null}
    },
    "actor": "job"
  }
}

Variable passing rules (uses "value", NOT "variable"):

  • {"task": "job", "value": "targetDevice"} → passes the parent’s targetDevice job variable to the child as deviceName
  • {"task": "static", "value": "validate"} → passes the literal string "validate"
  • {"task": "b2c3", "value": "return_data"} → passes a previous task’s output (preferred for runtime data)

WARNING — {task:"job"} refs in childJob variables add fields to inputSchema.required — same behavior as merge (see ### merge section). Only use {task:"job", value:"x"} for genuine workflow inputs. For runtime data produced by earlier tasks, use {task:"<taskId>", value:"<outVar>"} to reference the producing task directly.

WRONG for task output refs in childJob: {"task": "b2c3", "variable": "return_data"} — "variable" is for merge/evaluation only. In childJob, ALL refs (job, static, AND task output) use "value". Using "variable" causes undefined.indexOf() at job start time (P6.4.0+) — the workflow fails before any task runs.

Extracting single child output:

{
  "b2b2": {
    "name": "query",
    "variables": {
      "incoming": {
        "pass_on_null": false,
        "query": "taskStatus",
        "obj": "$var.a1a1.job_details"
      },
      "outgoing": {"return_data": "$var.job.childStatus"}
    }
  }
}

Query uses flat variable names — "taskStatus", NOT "variables.job.taskStatus".

If the query returns null even though the childJob succeeded — the $var form in obj is a stale incomingRefs entry (see the incomingRefs table under $var Resolution Rules — this is a caching issue, not a platform-version issue). Use the merge+taskRef workaround:

a1a1 (childJob) → c3c3 (merge: captures job_details via taskRef) → b2b2 (query: reads merged_object)
{
  "c3c3": {
    "name": "merge",
    "variables": {
      "incoming": {
        "data_to_merge": [
          {"task": "a1a1", "variable": "job_details"},
          {"task": "static", "value": {}}
        ]
      },
      "outgoing": {"merged_object": null}
    }
  },
  "b2b2": {
    "name": "query",
    "variables": {
      "incoming": {
        "pass_on_null": false,
        "query": "taskStatus",
        "obj": "$var.c3c3.merged_object"
      },
      "outgoing": {"return_data": "$var.job.childStatus"}
    }
  }
}

The static {} second item is required — merge needs at least 2 items in data_to_merge.

Mode B: Loop — one child per item in data_array

Each element in data_array becomes the child’s input variables for that iteration. Set variables: {} (empty).

Parent childJob task:

{
  "a1a1": {
    "name": "childJob",
    "canvasName": "childJob",
    "summary": "Run Child Per Device",
    "variables": {
      "incoming": {
        "task": "",
        "workflow": "My Child Workflow",
        "variables": {},
        "data_array": "$var.job.devices",
        "transformation": "",
        "loopType": "parallel"
      },
      "outgoing": {"job_details": null}
    },
    "actor": "job"
  }
}

Input: devices is an array of objects. Each object becomes one child’s variables:

{
  "devices": [
    {"deviceName": "IOS-CAT8KV-1", "action": "backup"},
    {"deviceName": "IOS-CAT8KV-2", "action": "check"},
    {"deviceName": "EOS-AWS-1", "action": "backup"}
  ]
}

Extracting loop output: Query "loop" to get the results array:

{
  "b2b2": {
    "name": "query",
    "variables": {
      "incoming": {
        "pass_on_null": false,
        "query": "loop",
        "obj": "$var.a1a1.job_details"
      },
      "outgoing": {"return_data": "$var.job.childResults"}
    }
  }
}

If the query returns null (platform-version-specific $var resolution issue), use the same merge+taskRef workaround described above (Mode A) — capture job_details via {"task": "a1a1", "variable": "job_details"} in merge, then query $var.c3c3.merged_object.

Loop element completeness — required fields must be in each element (not in variables).

The platform validates the child workflow’s inputSchema.required against each element’s keys only. Static variables set on the childJob task are NOT counted toward satisfying required fields. If your loop elements only contain per-iteration fields (e.g., subnet_name, subnet_cidr) but the child also requires shared fields (e.g., subscription_id, region), the validation fails before any iteration runs.

Fix — forEach enrichment pattern: enrich each element with the shared fields before the childJob loop, then set variables: {} on the childJob:

forEach (loop over elements) → merge (add shared fields to current_item) → arrayPush (append enriched element to new array)
                                                                                    ↓ (after forEach success)
childJob (data_array: enrichedArray, variables: {})
// forEach outgoing binds current_item to job var
{"outgoing": {"current_item": "$var.job.currentElement"}}

// merge combines current element + shared fields
{"data_to_merge": [
  {"task": "forEachId", "variable": "current_item"},
  {"key": "subscription_id", "value": {"task": "job", "variable": "subscription_id"}},
  {"key": "region", "value": {"task": "job", "variable": "region"}}
]}
// → $var.mergeId.merged_object is the enriched element

// arrayPush appends to accumulator
{"incoming": {"job_variable": "enrichedElements", "item_to_push": "$var.mergeId.merged_object"}}

// childJob uses the enriched array and no static variables
{"data_array": "$var.job.enrichedElements", "variables": {}, "loopType": "parallel"}

Loop output shape (each element is a flat spread of the child’s job variables):

[
  {"status": "complete", "childJobLoopIndex": 0, "deviceName": "IOS-CAT8KV-1", "action": "backup", "taskStatus": "success"},
  {"status": "complete", "childJobLoopIndex": 1, "deviceName": "IOS-CAT8KV-2", "action": "check", "taskStatus": "success"},
  {"status": "complete", "childJobLoopIndex": 2, "deviceName": "EOS-AWS-1", "action": "backup", "taskStatus": "success"}
]

Use "[**].taskStatus" in a query to extract one field from all iterations.

childJob checklist

  • actor is "job" (NOT "Pronghorn")
  • task is "" (empty string)
  • job_details outgoing is null
  • All incoming fields present — even unused ones: "data_array": "", "transformation": "", "loopType": ""
  • Variables use {"task","value"} NOT $var (single mode)
  • variables is {} when using data_array (loop mode)
  • Child workflow’s inputSchema.required matches what you’re passing
  • loopType: "" (single), "parallel" (simultaneous), "sequential" (one at a time)
  • If a downstream query of a childJob returns null: "obj": "$var.<childJobId>.job_details" is a stale incomingRefs entry, not a platform-version issue — use merge+taskRef workaround (see “Extracting single child output” above)

Building the child workflow

The child workflow must:

  1. Accept inputs via inputSchema that match what the parent passes
  2. Set output variables via newVariable or task outgoing → $var.job.x
  3. Handle errors internally (try-catch pattern) so it always completes:
task --success--> newVariable("taskStatus" = "success") -> workflow_end
task --error--> newVariable("taskStatus" = "error") -> workflow_end

The parent can then check taskStatus from job_details to decide what to do.


Projects

Method Endpoint Description
POST /automation-studio/projects/import Import a project (preferred — atomic)
POST /automation-studio/projects Create an empty project
GET /automation-studio/projects/{projectId} Get a project
PATCH /automation-studio/projects/{projectId} Update a project
DELETE /automation-studio/projects/{id} Delete a project
GET /automation-studio/projects/{id}/export Export project as JSON
POST /automation-studio/projects/{projectId}/components/add Add components (legacy)
DELETE /automation-studio/projects/{projectId}/components/{componentId} Remove component

Preferred: Import a project (atomic — all assets in one call)

Before writing the first line of a project/workflow import payload, open assets/helpers/create/import-project.json and start from that scaffold. Do not build the {project: {components: [...]}} shape from scratch in memory, even if you’re confident in it — the wrapper shape (project-level vs. component-level metadata, which fields need _id vs. which don’t, the exact created_by shape difference between project and workflow) is easy to get subtly wrong in ways that produce contradictory-looking “additional properties” / “required property” errors that seem like they’re about unrelated fields (e.g., groups).

If the platform’s import mechanism you’re using is a file-upload UI dialog (not a REST call you make directly), the payload shape may differ from the POST /automation-studio/projects/import API documented below. Always call the documented API endpoint yourself via curl/HTTP client rather than asking a human to paste/upload a file through a UI — you have the credentials and the ability to make the call directly; routing through a human as a manual courier for a schema you haven’t verified multiplies the round-trip cost of every guess. If a UI-only import path is genuinely the only option available, treat its error messages with the same Repeat-Failure Circuit Breaker discipline as any other validation error — don’t iterate blindly.

Always use import instead of create + add components. Import creates the project with all workflows, templates, and MOP templates inside it in a single atomic call. No intermediate state, no broken childJob refs, no project-locking issues.

⚠️ Data-loss warning: importing against an EXISTING project _id replaces that project’s entire components array — it does not merge or add to it. If you’re reusing a helper asset (e.g., cloning assets/helpers/assets/itential-platform-email.json) to add a NEW capability to a project that already has other components in it, and you set the payload’s _id to that existing project’s ID, every component not listed in your payload will be silently deleted — including workflows built and verified in a prior session. The platform works this way by design, answers HTTP 200 with failedComponents: [], and does not report what was added, kept or removed. Before every import against an existing project _id, follow the procedure below — do not assume the call is additive just because you’re only trying to add one new thing.

Before importing over an existing project, compare and confirm.

  1. Check whether the project _id already exists: GET /automation-studio/projects/{id}. If it returns 404, this is a new project and you can import.
  2. If it exists, export it: GET /automation-studio/projects/{id}/export, and build your payload from the export’s data object — it is already in import format (clean names, user objects, no encodingVersion), so send it as {"project": <data>} with your changes applied. Don’t use the project GET as the source (it returns only a pointer per component) or a stored project-import.json (it may be out of date). Also record the project’s current members (owners and groups) from the project GET: an import sets the service account as owner (Rule 11a) and may drop the existing members, and the membership PATCH is a full replacement, so you need the list to restore them.
  3. Compare your payload with the live project’s components (name and type; strip the @<projectId>: prefix). Make three lists: kept (in both), added (only in your payload) and removed (only in the live project).
  4. Tell the user what will happen, in plain words: “This import replaces project X. It will keep A, B; add C; and delete D, E.” Don’t import until the user says yes.
  5. If the user only wanted to add or change something and the removed list is not empty, do not import. Rebuild the payload from the export so nothing is removed, or use conflictMode: "insert-new" to create a separate copy and leave the original untouched. If you use insert-new, check the copy’s childJob workflow refs afterward — they were pre-wired with the original project’s @{projectId}: prefix, and the platform may not rewrite them — and fix any that still point at the original project.
  6. Make the membership PATCH your very next call after the import (Rule 11a), restoring the members you recorded in step 2. Then read the project again, check that its components match the lists you showed the user, and report any difference.

conflictMode (import option): "overwrite" is the default and replaces the existing project; "insert-new" leaves the existing project alone and creates a separate new project from the file. Do not combine insert-new with assignNewReferences: true (one test returned HTTP 500; cause unknown). A project can depend on something outside it (for example an LCM resource model, listed in the export’s referencedComponentHashes); if the target platform lacks it, import fails reference validation unless skipReferenceValidation: true is sent.

Template document IDs are re-created on every import, so do not rely on a stored template ID after a re-import.

POST /automation-studio/projects/import

Build all assets locally first, then import everything at once:

{
  "project": {
    "_id": "24-char-hex-mongodb-objectid",
    "iid": 1,
    "name": "My Project",
    "description": "...",
    "thumbnail": "",
    "backgroundColor": "#FFFFFF",
    "components": [
      {
        "iid": 1,
        "type": "workflow",
        "reference": "uuid-of-workflow",
        "folder": "/",
        "document": { "...full workflow object..." }
      },
      {
        "iid": 2,
        "type": "mopCommandTemplate",
        "reference": "@projectId: Template Name",
        "folder": "/",
        "document": { "...full MOP object..." }
      }
    ],
    "created": "2026-03-13T00:00:00.000Z",
    "createdBy": {"_id": "000000000000000000000000", "provenance": "CloudAAA", "username": "admin@itential"},
    "lastUpdated": "2026-03-13T00:00:00.000Z",
    "lastUpdatedBy": {"_id": "000000000000000000000000", "provenance": "CloudAAA", "username": "admin@itential"}
  }
}

Import format rules (different from create; a project export already follows them):

Field Import format Notes
encodingVersion OMIT from workflow documents Causes silent component failure if included
created_by (workflow) {username, provenance, firstname, inactive, sso} — NO _id Different from project-level createdBy
createdBy (project) {_id, username, provenance} — HAS _id Different from workflow-level
_id (project) Pre-compute 24-char hex string So childJob refs can use @{projectId}:
Workflow name Clean names — no prefix Import adds @projectId: automatically
childJob workflow Must include @{projectId}: prefix Pre-wire using the same _id
reference (workflow) UUID string Becomes the workflow’s uuid
reference (MOP) @{projectId}: Template Name String reference
reference (template, transformation, JSON form) 24-character hex id, not a uuid A uuid here fails the import with Argument passed in must be a string of 12 bytes or a string of 24 hex characters, which names nothing
iid (components) Sequential integers starting at 1 Incrementing ID

NEVER feed a GET /automation-studio/workflows/detailed/{name} read-back document into projects/import unmodified — it crashes the import and, against an EXISTING project _id, wipes its components. The platform mutates workflow documents after import/save: created_by/last_updated_by become plain account-ID strings (import requires the object shape above), outputSchema is expanded with platform-generated fields, and encodingVersion/_id/errors/warnings/namespace appear. Importing such a document fails every component with "WorkflowBuilder stopped during execution" — and because a re-import is a destructive full-replace, the project is left with components: [] (confirmed data loss, recovered only by re-importing the ORIGINAL known-good documents). This warning is about that one endpoint (a single workflow read-back), not about everything read from the platform. The project export (GET /automation-studio/projects/{id}/export) re-imports cleanly: its data object is already in import format (component names without the @projectId: prefix, user objects, no encodingVersion), and real projects with transformations and scenarios survive an export and import round trip (confirmed in testing). To re-import a workflow read back from workflows/detailed: reuse the original import documents, or normalize the read-back first (restore object-shaped created_by/last_updated_by, strip encodingVersion, _id, errors, warnings, namespace, and strip the @projectId: name prefix). Either way, follow the compare-and-confirm procedure above before importing over an existing project.

Response:

{
  "message": "Successfully imported project",
  "data": {"_id": "...", "name": "...", "components": [...]},
  "metadata": {"failedComponents": []}
}

Check metadata.failedComponents — empty array means success.

Why import instead of create + move

Problem Create + move Import
childJob refs Break on move — manual fix needed Pre-wired with @projectId: — just work
Project locking Race conditions during move Single atomic call
Intermediate state Workflows exist outside project Never
API calls Create + create each asset + move + fix refs One POST
Reproducibility Hard to replay project-import.json is the artifact

If a project reference you were relying on turns out to be missing or stale (e.g., GET /automation-studio/projects/{id} returns "Project not found" for a project you thought you’d already created), do not respond by creating a brand-new project and moving a pre-existing standalone workflow into it via components/add. That’s the create-then-move pattern above, just arrived at while recovering from an unrelated error instead of choosing it deliberately — it’s still the discouraged pattern. Rebuild the intended end state atomically via projects/import instead, even if that means recreating a workflow that technically already exists standalone elsewhere.

Legacy: Create + add components (avoid if possible)

Only use this for adding a single asset to an existing project after initial import.

POST /automation-studio/projects/{projectId}/components/add
{
  "components": [
    {"type": "workflow", "reference": "uuid-...", "folder": "/"}
  ],
  "mode": "move"
}

Warning: Both move and copy rename assets with @projectId: prefix but do NOT update internal references (childJob workflow fields, template names). You must fix these manually.

A bare {type, reference} component does NOT embed the workflow’s tasks/transitions — it stores a pointer to a separately-existing workflow document. If you fetch the project afterward and the component has no document field (or an empty one), this is why — not a save failure. This is the most common reason components/add looks like it silently failed, and it’s the root cause of the “create-then-move” anti-pattern above: reaching for components/add expecting it to carry the full document, discovering it didn’t, deleting and retrying. If you need the full document inside the project, use atomic projects/import with the document nested under each component from the start instead — don’t retry components/add variations expecting different behavior.

Component types: workflow, template, transformation, jsonForm, mopCommandTemplate, mopAnalyticTemplate

components/add called more than once on the same project can corrupt iid tracking. Confirmed on a project created via empty POST /automation-studio/projects (not import): a first components/add call (e.g. 5 components) gets assigned sequential iids, but the project’s own componentIidIndex can come back reporting a value lower than the actual max iid just used. A second components/add call trusts that stale index as its starting point and assigns new iids that collide with ones from the first call — two different components end up sharing the same iid, silently corrupting anything that references components by iid (notably folders[].children[].iid groupings).

Diagnostic sign: after adding components in more than one call, GET the project and check components[].iid for duplicates across different types.

PATCH does not fix this. Sending a corrected components/folders array with de-duplicated, hand-assigned iids via PATCH /automation-studio/projects/{projectId} returns 200, but the platform recalculates/discards your custom iid values server-side — a GET right after still shows the same collision. Don’t try to patch your way out of it.

Fix — pick one: (a) batch every component you’re adding into a single components/add call (confirmed reliable — produces clean sequential iids), or (b) rebuild the project from scratch via projects/import, embedding each component’s full document directly rather than moving in pre-existing standalone assets.

Update membership (full replacement)

Before patching, always ask the engineer: “Who else should have access to this project? (usernames or group names)”

Do not auto-discover or assume groups. Wait for the answer, resolve each name to a reference ID by scanning existing projects, then PATCH.

PATCH /automation-studio/projects/{projectId}

Use the helper: assets/helpers/update/update-project-members.json

Include ALL members in every PATCH — this is a full replacement. Omitting an existing member removes them.

If regular PATCH /automation-studio/projects/{projectId} returns 403 for the service account (no GBAC role on the project yet), the bypass endpoint PATCH /automation-studio/admin/projects/{projectId} will succeed — but it silently drops any component not listed in the SAME request body, even when the request only intends to change members. Confirmed live: PATCHing {"members": [...]} alone (no components key at all) removed a real mopCommandTemplate component from the project — the underlying asset was deleted outright, not just unlinked, and had to be recreated from a backup. Always GET the project immediately after any admin-endpoint PATCH and diff the components array against what it was before; if anything is missing, recreate it (via the asset’s own create endpoint, e.g. POST /mop/createTemplate) and re-link with components/add.

To resolve a username or group name to a reference ID, scan existing projects:

for pid in $(curl -s "$BASE/automation-studio/projects?limit=100" \
  -H "Authorization: Bearer $TOKEN" | jq -r '.data[]._id'); do
  curl -s "$BASE/automation-studio/projects/$pid" \
    -H "Authorization: Bearer $TOKEN" \
    | jq -r '.data.members[]? | [.type, .reference, (.username // .name)] | @tsv'
done | sort -u

If a name cannot be resolved, ask the engineer for the reference ID — do not guess.

Resolve membership references from spec

Your very next tool call after a successful projects/import must be the membership PATCH below — before verifying the import, before building the next component, before anything else. Import sets the OAuth service account as sole project owner, not the UI user from the spec. If your next action isn’t this PATCH, you’ve skipped it — reading this warning is not the same as having acted on it. This runs in Phase 3 (Import), not Phase 6 (Deliver).

There is no user/group lookup API on the Itential platform. The only way to resolve a username (e.g., joksan.flores@itential.com) or group name (e.g., solutions-engineers) to a platform reference ID is by scanning existing projects’ members.

Step 1: Build a membership lookup table.

The list endpoint (GET /automation-studio/projects?limit=50) does NOT include username/name on member objects — only individual GET /automation-studio/projects/{id} calls do. Scan all projects to build the lookup:

# Get all project IDs
PROJECT_IDS=$(curl -s -H "Authorization: Bearer $TOKEN" \
  "$PLATFORM_URL/automation-studio/projects?limit=100" \
  | jq -r '.data[]._id')

# Build lookup table from individual GETs
> {use-case}/membership-lookup.txt
for pid in $PROJECT_IDS; do
  curl -s -H "Authorization: Bearer $TOKEN" \
    "$PLATFORM_URL/automation-studio/projects/$pid" \
    | jq -r '.data.members[]? | [.type, .reference, (.username // .name), .provenance] | @tsv'
done | sort -u >> {use-case}/membership-lookup.txt

Output format (TSV): type reference username/name provenance

Step 2: Match spec members to references.

For each member in the spec’s Project Membership table, find their reference ID in membership-lookup.txt:

grep "joksan.flores@itential.com" {use-case}/membership-lookup.txt
# → account  699a67bb...  joksan.flores@itential.com  CloudAAA

Step 3: Cache the result — don’t re-scan for the same person next time. Append resolved {username: accountId} pairs to use-case-memory.md (or a shared membership-lookup.txt if working across multiple use-cases). Re-scanning every project from scratch each time the same engineer needs to be added as owner on a new use-case is a wasted 50-100+ API calls when the answer was already resolved once. If use-case-memory.md doesn’t exist yet for this engagement (e.g., in freestyle/explore-mode work with no formal spec), create it anyway — see the Directory Layout section’s “living context” note.

Step 3: PATCH membership immediately after import.

PATCH /automation-studio/projects/{projectId}
{
  "members": [
    {"type": "account", "role": "owner", "reference": "699a67bb..."},
    {"type": "group", "role": "editor", "reference": "67c859..."}
  ]
}

If a username or group cannot be resolved from the lookup table, stop and ask the engineer. Do not guess reference IDs or skip members.

Baseline members (when no spec membership is defined): If there is no Project Membership table in the spec, or when doing a freeform build/import outside the spec lifecycle, ask the engineer: “Which user accounts or groups should have access to this project?” — do not assume or skip. Once you have the names, resolve them via the lookup table above and PATCH immediately. Without this step the engineer will be locked out of the project in the IAP UI. See #63

Project Thumbnail

Operation Endpoint
Set PUT /automation-studio/projects/{id}/thumbnail — body: {"imageData": "<data-URI>", "backgroundColor": "<hex>"}
Get GET /automation-studio/projects/{id}/thumbnail — returns {"data": {"image": "<data-URI>", "backgroundColor": "<hex>"}}

imageData must be a full data URI — not raw base64. Passing raw base64 without the data:image/png;base64, prefix returns HTTP 200 and stores the value, but the UI renders a black/blank image with no error.

data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAA...

Build the data URI in Python:

import base64, io
buf = io.BytesIO()
img.save(buf, format='PNG')
data_uri = f"data:image/png;base64,{base64.b64encode(buf.getvalue()).decode()}"
  • Optimal dimensions: 330 × 100 px — matches the project card aspect ratio in Automation Studio
  • Accepted formats: jpg, jpeg, png — max 1000 KB
  • backgroundColor (hex, e.g. "#1B2A4A") sets the card background color visible before the image loads

JSON Forms

JSON Forms have their own dedicated skill — itential-json-forms. See that skill for the form structure (struct / schema / uiSchema / bindingSchema), the static-enum vs. REST-bound vs. cascading dropdown (aka field dependency) patterns, the full API reference (including the bulk-only DELETE), and the manual-trigger wiring (legacyWrapper: false).

Helper templates for forms still live under assets/helpers/:

  • assets/helpers/create/create-json-form.json — static-enum dropdowns
  • assets/helpers/create/create-json-form-rest-bound.json — REST-bound or cascading dropdowns
  • assets/helpers/create/create-json-form-array-field.json — array-of-objects field (every field, including nested ones, needs a nodeId)

Operations Manager (Automations & Triggers)

Method Endpoint Description
POST /operations-manager/automations Create an automation
GET /operations-manager/automations List automations
POST /operations-manager/triggers Create a trigger
PATCH /operations-manager/triggers/{id} Update a trigger
GET /operations-manager/triggers List triggers

Create a Manual Trigger with JSON Form

This is a two-step process: create the automation, then create a manual trigger that binds to it.

Use the helper template: assets/helpers/create/create-ops-manager-automation.json

Critical: legacyWrapper must be false. When creating a manual trigger with a JSON form, set legacyWrapper: false. The default is true, which wraps form field values under formData, breaking the mapping to workflow job variables. With legacyWrapper: false, form field values map directly to workflow input variables by name.

Required trigger fields: name, type ("manual"), enabled, actionType ("automations"), actionId, formId, legacyWrapper


Task Discovery

Pull Task Catalog

{use-case}/tasks.json should already exist — pulled by /solution-arch-agent or /explore during feasibility. Do not re-pull if the file exists. If missing, fetch it:

GET /workflow_builder/tasks/list → save to {use-case}/tasks.json
GET /automation-studio/apps/list → save to {use-case}/apps.json

Search locally:

grep -i "template" {use-case}/tasks.json
jq '.[] | select(.app == "ConfigurationManager") | .name' {use-case}/tasks.json

Look up task wiring in asset projects first

Before fetching schemas from the API, check if an asset project already has the task wired up. If it does, you get the exact field structure for free — no API call needed.

# Does any asset project use this task? Find it by task name:
grep -rl '"name": "TASK_NAME"' assets/helpers/assets/

# Extract the wired task from the matching project:
jq '[.components[].document.tasks // {} | to_entries[] | select(.value.name == "TASK_NAME")] | first | .value' \
  assets/helpers/assets/MATCHING_FILE.json

# See which tasks a specific workflow uses:
jq '[.components[] | select(.type=="workflow") | select(.document.name | test("WORKFLOW"; "i"))] | first | .document.tasks | to_entries[] | {id:.key, name:.value.name, app:.value.app}' \
  assets/helpers/assets/MATCHING_FILE.json

Asset project → best match by task type:

Task Best asset to check
ServiceNow adapter tasks assets/helpers/assets/vendor-servicenow.json
Infoblox / DNS / IPAM tasks assets/helpers/assets/vendor-infoblox-nios-ddi.json
NetBox tasks assets/helpers/assets/vendor-netbox.json
itential_cli, RunCommandTemplate, MOP tasks assets/helpers/assets/itential-platform-configuration-management.json, assets/helpers/assets/vendor-cisco-ios.json
transformation (JST) assets/helpers/assets/vendor-netbox.json, assets/helpers/assets/itential-platform-data-manipulation.json
childJob, evaluation, query, newVariable any vendor project
LCM action workflow tasks assets/helpers/assets/lcm/lcm-vxlan-fabric-services-project.json

Get Full Task Schemas (only if not found in assets)

Single task:

GET /automation-studio/locations/{location}/packages/{pckg}/tasks/{method}?dereferenceSchemas=true

Multiple tasks:

POST /automation-studio/multipleTaskDetails?dereferenceSchemas=true
{
  "inputsArray": [
    {"location": "Application", "pckg": "WorkFlowEngine", "method": "query"},
    {"location": "Adapter", "pckg": "Servicenow", "method": "createChangeRequest"}
  ]
}

Mapping from tasks.json → schema endpoint:

tasks.json field Maps to
location (Application/Adapter) {location}
app (e.g., TemplateBuilder) {pckg}
name (e.g., renderJinjaTemplate) {method}

IMPORTANT: The pckg value must come from apps.json, NOT tasks.json. The names can differ (e.g., tasks.json says ServiceNow but apps.json says Servicenow).

Before fetching schemas:

  1. Search asset projects (above) — if found, use the wired example directly
  2. Check if {use-case}/task-schemas.json exists — search it next
  3. Only call multipleTaskDetails for tasks not found in either place
  4. After fetching, append to {use-case}/task-schemas.json

Adapter list/search tasks (e.g. ipam_prefixes_list, ipam_ip_addresses_list) require every declared query parameter present in incoming, even the ones you don’t care about — as an empty string. Wiring only the parameters you actually want to set (e.g. just adapter_id and one filter) is enough to pass workflows/validate, but crashes jobs/start with the same opaque, task-unattributed "Cannot convert undefined or null to object" covered above. Copy the FULL incoming block from a real wired example in the asset projects rather than hand-picking fields, and only overwrite the specific ones you need.

Check each parameter’s declared type before assigning it a plain string — some are array-typed even when sibling parameters with similar names are plain strings. Confirmed with NetBox’s ipam_ip_addresses_list: nearly every filter parameter is declared type: array in the dereferenced schema (via multipleTaskDetails?dereferenceSchemas=true), including ones like parent that read like a single-value filter. Passing a plain string ("parent": "100.64.1.0/24") for an array-typed parameter fails at jobs/start with a swagger-client error, "Could not parse parameter value string as JSON Object or JSON Array" — pass a real JSON array instead ("parent": ["100.64.1.0/24"]).

A _bulk_destroy-style adapter task that issues DELETE with a request body will fail at the gateway/proxy layer, not the adapter or the target API — confirmed with NetBox’s ipam_ip_addresses_bulk_destroy (requestBodyPayload on a DELETE): the call never reaches NetBox, failing instead with "error receiving task event: rpc error: ... invalid spec: HTTP method HTTP_METHOD_DELETE does not support request body". This is a hard gateway-proxy limitation, not something fixable via task configuration — for gateway-routed adapters, avoid any DELETE-with-body bulk operation and loop single-item _destroy calls instead (see ### forEach above for the loop pattern; state: "loop", not childJob, since the target of the loop body here is an adapter task rather than another workflow).

nodeLocation Spacing Convention

Workflows are laid out top-to-bottom (vertical) by default — this is the Itential best practice for readability and consistency, and matches the conventions used in the platform’s working examples. Use horizontal only when the engineer explicitly asks for it.

Vertical Layout (default)

Rule Value
Sequential tasks (y-delta) +108px
Fork branch offset from spine (x-delta) ±264px
Spine x a constant column (e.g. x=600)

Clean canvas principles:

  • The spine is a constant x — non-forked sequences (start, single-thread tasks, end, convergence points) sit on it.
  • Forks split off the spine — at a fork point, both outgoing branches leave the spine column. Place one at spine - 264 and the other at spine + 264. The spine column stays empty between the fork and the convergence point so transition lines don’t cross task nodes. Direction (which branch goes left vs. right) is the engineer’s call — pick whatever keeps the picture clean.
  • Branches stay in their own column until they converge.
  • Convergence tasks (workflow_end, merges, error sinks) return to the spine x.
  • Tight y-spacing — the canvas grid is dense; ~108px between sequential rows reads well. Don’t pad to +250 or +360.
  • Preserve Studio-arranged positions — if an engineer has arranged a workflow in Automation Studio, treat its nodeLocation values as authoritative. Always read from the live export before reimporting. Never recalculate positions from scratch on a workflow that has already been arranged.

Example — fork with a shared error handler (same pattern as ServiceNow “Create Change Request” in assets/helpers/assets/vendor-servicenow.json):

workflow_start                        (x=600, y=200)
e1a1 merge                            (x=600, y=312)
a1b2 createCR  ── fork point ──       (x=600, y=420)
b2c3 query   [success branch]         (x=336, y=540)
c3d4 updateCR [success branch]        (x=336, y=636)   ef01 newVar [shared error handler]  (x=864, y=636)
workflow_end                          (x=600, y=804)

For a childJob phase with query + evaluation (single-thread, no fork → all on spine):

y=312  — childJob     (x=600)
y=420  — query        (x=600)   ← extracts taskStatus from job_details
y=528  — evaluation   (x=600)

Horizontal Layout (only when requested)

If the engineer explicitly asks for horizontal, swap x and y throughout: phases advance on x, fork branches offset on y, spine becomes a constant y row. Same magnitudes, opposite axes.

Verifying Layout After Build — Don’t Just Eyeball the Checklist

The spine/fork/convergence convention above is easy to violate without noticing, because each task’s nodeLocation is usually set incrementally as it’s created — relative to whatever was placed immediately before it, not to the graph as a whole. A workflow can be 100% correctly wired (every transition right, every $var resolving) and still render with failure/error-branch tasks scattered at inconsistent x-offsets and heights, each needing a long diagonal line back to wherever the workflow actually ends. That’s a real, observed failure mode — not hypothetical — and the checklist bullets above only catch it if you deliberately re-read every nodeLocation against the convention, which is easy to skip once the workflow otherwise looks “done.”

Verify mechanically instead of by eye. After wiring all tasks and transitions, before calling Build complete, run something like this against the built workflow’s tasks/transitions:

import json

wf = json.load(open("workflow.json"))  # or wf["items"][0] from a GET
tasks, transitions = wf["tasks"], wf["transitions"]

xs = [t["nodeLocation"]["x"] for tid, t in tasks.items() if tid not in ("workflow_start", "workflow_end")]
spine = max(set(xs), key=xs.count)  # most common x = the spine

violations = []
for tid, t in tasks.items():
    if tid in ("workflow_start", "workflow_end"):
        continue
    x = t["nodeLocation"]["x"]
    if x != spine and abs(x - spine) != 264:
        violations.append(f"{tid} ({t.get('summary') or t.get('name')}): x={x}, not spine ({spine}) or spine±264")

# Flag any task whose incoming transitions come from tasks with wildly different y —
# a large y-gap into a task usually means an earlier removal/reorder left a stale position.
# Skip revert transitions entirely: they're supposed to go backward (retry loops), so a
# negative y-delta there is correct wiring, not a layout violation.
for src, dsts in transitions.items():
    if src not in tasks:
        continue
    for dst, edge in dsts.items():
        if dst not in tasks or edge.get("type") == "revert":
            continue
        dy = tasks[dst]["nodeLocation"]["y"] - tasks[src]["nodeLocation"]["y"]
        if dy < 0 or dy > 200:
            violations.append(f"{src} -> {dst}: y-delta={dy} (expect ~108, or a deliberate fork/convergence jump)")

if violations:
    print("Layout violations found — fix nodeLocation before considering Build done:")
    for v in violations:
        print(" -", v)
else:
    print("Layout OK: spine =", spine)

Re-run this after every PUT that adds, removes, or rewires a task — not just once at the end. Removing a task (e.g. an error-handling branch that got redesigned out) is a common way to leave a gap that stretches every task after it, which this catches immediately instead of leaving it for the engineer to notice on the canvas later.


Workflows

Workflow Structure

POST /automation-studio/automations

Body wraps the workflow in {"automation": {...}}:

{
  "automation": {
    "name": "My Workflow",
    "description": "Does something useful",
    "type": "automation",
    "canvasVersion": 3,
    "encodingVersion": 1,
    "font_size": 12,
    "tasks": {
      "workflow_start": {
        "name": "workflow_start",
        "groups": [],
        "nodeLocation": {"x": 600, "y": 200}
      },
      "a1b2": {
        "name": "query",
        "canvasName": "query",
        "summary": "Extract Data",
        "description": "Extracts field from response",
        "location": "Application",
        "locationType": null,
        "app": "WorkFlowEngine",
        "type": "operation",
        "displayName": "WorkFlowEngine",
        "variables": {
          "incoming": {
            "pass_on_null": false,
            "query": "hostname",
            "obj": "$var.job.deviceData"
          },
          "outgoing": {
            "return_data": "$var.job.deviceName"
          },
          "error": "",
          "decorators": []
        },
        "groups": [],
        "actor": "Pronghorn",
        "scheduled": false,
        "nodeLocation": {"x": 600, "y": 312}
      },
      "workflow_end": {
        "name": "workflow_end",
        "groups": [],
        "nodeLocation": {"x": 600, "y": 420}
      }
    },
    "transitions": {
      "workflow_start": {
        "a1b2": {"type": "standard", "state": "success"}
      },
      "a1b2": {
        "workflow_end": {"type": "standard", "state": "success"}
      },
      "workflow_end": {}
    },
    "groups": [],
    "inputSchema": {
      "type": "object",
      "properties": {
        "deviceData": {"title": "deviceData", "type": "object"}
      },
      "required": ["deviceData"]
    },
    "outputSchema": {
      "type": "object",
      "properties": {
        "deviceName": {"title": "deviceName", "type": "string"}
      }
    }
  }
}

Scenarios. A workflow can carry scenarios, the mock data used in the canvas debug mode. Each scenario needs name and input, and input must be an object ({} when empty), never a string. A string input fails the validator with Schema violation at /scenarios/0/input: must be object.

Update a workflow:

PUT /automation-studio/automations/{id}
{"update": { ...same structure as automation object... }}

Project-scoped name required on PUT. If the workflow belongs to a project, the name field in the update body must include the @<projectId>: prefix — even if the workflow was created without it:

{"update": {"name": "@69f10abc: My Workflow", "tasks": {...}, "transitions": {...}}}

Sending the bare name ("name": "My Workflow") returns {"error": {"message": "Name must begin with '@projectId: '"}}.

Asymmetry: workflow CREATE (POST /automation-studio/automations) does NOT require the prefix — the platform applies it when the workflow is added to a project. But PUT-update always requires it for project-member workflows.

Always read the workflow before updating (GET /automation-studio/workflows/detailed/{name} or export the project) to get the current scoped name. See Rule 24 and issue #55. Build the update body from the workflow you just read, so fields such as canvasVersion are kept.

Task Fields

Field Application Tasks Adapter Tasks
name Method name from tasks.json Method name from tasks.json
canvasName From tasks.json canvasName field (may differ from name: arrayPush→push) Same
location "Application" "Adapter"
locationType null Same as app
app App name (e.g., WorkFlowEngine) From apps.json (NOT tasks.json)
type "automatic" or "operation" — read from tasks.json .type, do not guess
actor "Pronghorn" "Pronghorn"
displayName App name May differ from app

Adapter tasks also require adapter_id in incoming variables — the adapter instance name from health/adapters.

Task Access Control (groups)

The groups field on a task definition is task-level GBAC — group-based access control that restricts which IAP groups can see, claim, and complete a manual task in the Job Inbox.

Field Type Meaning
groups (plural) string[] GBAC. Each entry is a group’s MongoDB _id (24-char hex). Empty [] means no task-level restriction.
group (singular, optional) string Canvas display category (e.g., "Tools", "JsonForms"). Set by the Studio canvas. NOT access control — easy to confuse with groups.
{
  "name": "ViewData",
  "type": "manual",
  "app": "WorkFlowEngine",
  "view": "/workflow_engine/task/ViewData",
  ...
  "groups": ["69e65b4189b39131a9b8cce1"]
}

Look up group IDs:

  • GET /authorization/groups — list groups (each has _id and name)
  • GET /authorization/groups/<id> — resolve a single group

Two GBAC scopes — both use the same string[] shape (group _ids) but apply at different levels:

  • Per-task groups (on the task definition, sibling of name/app/type) — gates access to a single manual task.
  • Top-level workflow groups (sibling of tasks/transitions at the workflow level) — gates access to the workflow as a whole.

Tasks of any type can carry groups, but only type: "manual" tasks surface in the Job Inbox where GBAC actually gates user access. Leave it as [] on automatic tasks unless platform-specific docs say otherwise.

Edge cases not yet documented — verify on your platform before relying on:

  • Semantics with multiple group IDs in the array (likely OR — any-of — but unverified)
  • Interaction between task-level and workflow-level groups (additive vs. override)
  • Whether groups accepts a $var job-variable for dynamic group resolution (almost certainly no — design-time only — but worth confirming)

Task IDs

Task IDs must be hex-only: [0-9a-f]{1,4}, plus the two reserved names workflow_start and workflow_end. Non-hex IDs (e.g., apush) cause $var references to silently fail. This includes the start/end tasks themselves — do not key them as "start"/"end" (a natural-sounding but invalid shortcut); the reserved keys are the literal strings workflow_start and workflow_end, not any descriptive substitute. Using "start"/"end" (or any other non-hex, non-reserved key) produces a generic "must NOT have additional properties" import error that looks unrelated to task naming — if you see that error and your task keys include anything other than hex IDs or the two reserved names, fix the keys first before investigating any other field.

Naming a new or inserted task: give it a new, unused ID of 1 to 4 hex characters (0-9, a-f). Do not make up readable or numbered IDs with a prefix (p100, step2), and do not build the ID from a neighbor’s by adding a letter or suffix (c002b, a1b2_new). A key longer than 4 characters, or with a non-hex character, is rejected on import with a generic must NOT have additional properties error. IDs do not need to sort in task order.

Transitions

"transitions": {
  "workflow_start": {
    "a1b2": {"type": "standard", "state": "success"}
  },
  "a1b2": {
    "c3d4": {"type": "standard", "state": "success"},
    "e5f6": {"type": "standard", "state": "error"}
  },
  "c3d4": {
    "workflow_end": {"type": "standard", "state": "success"}
  },
  "e5f6": {
    "workflow_end": {"type": "standard", "state": "success"}
  },
  "workflow_end": {}
}

Transition states:

  • success — task completed without error (all tasks)
  • error — task encountered errors (all tasks)
  • failure — evaluation didn’t match or query returned undefined (evaluation/query only)
  • loop — forEach loop iteration (forEach only)

Transition types:

  • standard — moves forward
  • revert — moves backward to a previous task (retry loops)

MANDATORY: Every adapter/external task needs an error transition. Without one, errors cause “Job has no available transitions” and the job gets stuck forever. POST /workflow_engine/workflows/validate does not check for this — it only flags a task missing a success path, never a missing error path. A workflow with no error transitions anywhere can still return isValid: true. Do not skip this check because validate passed.

JSON duplicate key problem: If both success and error need to go to workflow_end, you can’t use workflow_end as a key twice. Route error to an intermediate task (e.g., newVariable to set error status), then route that to workflow_end:

"transitions": {
  "a1b2": {
    "c3d4": {"type": "standard", "state": "success"},
    "e5f6": {"type": "standard", "state": "error"}
  },
  "e5f6": {
    "workflow_end": {"type": "standard", "state": "success"}
  }
}

Create Response Shape

Both workflow and template creation return {created, edit} — NOT {message, data, metadata}:

{
  "created": {"_id": "...", "name": "..."},
  "edit": "/automation-studio/#/edit?..."
}

$var Resolution Rules

$var only resolves as direct top-level incoming variable values:

Wiring Works? Why
"deviceName": "$var.job.x" Yes Direct top-level value
"variables": {"key": "$var.job.x"} NO Nested inside object
"body": {"data": "$var.job.x"} NO Nested — stored as literal string

Workaround: Use merge, makeData, or query to build the nested object, then reference the task’s output with $var.taskId.merged_object.

Exception — for the object-typed fields in the exception table below, don’t just apply this workaround; the right shape depends on how Studio edits each field (Platform 6.5.2, 2026-10):

  • runCode.data (keyed rows, each with its own Static/Job/Task source): one $var per key, e.g. "data": {"deviceName": "$var.job.deviceName", "nb": "$var.b1b3.response"}. Studio shows these as bound rows. A whole-object string ref ("data": "$var.a1a1.merged_object") runs, but Studio renders it as one Static row per character (data.0 = $, data.1 = v, …) and “Add Input Data Key” throws TypeError: Cannot use 'in' operator to search for 'input_N' in $var.... Editing a row there (changing its value or source, or deleting one) and saving writes the character keys into the stored data ({"0": "$", "1": "v", ...}), so the script receives those one-character keys instead of the object, and the job still completes with no error. Opening and saving without touching the data panel leaves the string intact. No validator flags it.
  • runAgent.inputs (rows built from the agent’s inputSchema, each with its own Static/Job/Task source): one $var per key, named exactly as the agent’s input schema, e.g. "inputs": {"first_string": "$var.job.first", "second_string": "$var.job.second"}. Studio shows these as bound rows. A whole-object string ref ("inputs": "$var.x.merged_object") is not shown in the panel at all; editing any row there and saving stores one key per character beside the edited key ({"0": "$", "1": "v", ..., "first_string": "abc"}), loses the ref and the other inputs, and the session then fails with must have required property ....
  • runService.params (depends on whether the service declares a parameter schema; check service_metadata.decorator in GET /gateway_manager/v1/services):
    • Service with a declared schema (e.g. run-command): Studio shows a Key/Value Form with one row per schema property, each with its own Static/Job/Task source. Write one $var per key, "params": {"command": "$var.job.command"}; Studio shows it as bound rows. A whole-object string ref is not shown in that form (empty Static rows), and editing a row then saving replaces it with the typed values ({"command": "abc123"}), silently dropping the ref.
    • Service without a schema (e.g. a bare python-script): Studio shows params as one object with a single Static/Job/Task source. Wire it to a merge output, "params": "$var.<mergeId>.merged_object"; Studio shows Variable Source = Task. Per-key $var strings in that object resolve at runtime, but Studio shows them as a Static JSON value with literal text, so don’t author them.
  • transformation.variableMap: one $var per key. A whole-object string ref makes jobs/start fail with Cannot assign to read only property '0' of string '$var...'. Studio shows the per-key form as bound rows, each with its own Static/Job/Task source.

Task ID validation: $var.taskId.x only resolves when taskId matches [0-9a-f]{1,4}. Non-hex IDs silently fail.

Keep variables to a minimum. Read a value from the task that produced it — $var.<taskId>.<outVar>, or a field of it with Enable Query ($var.<taskId>.result#/stdout_json/<field>). Don’t add newVariable or query tasks just to hold or extract a value, and put data shaping in one runCode task rather than a chain of utility tasks. Create a job variable ($var.job.x) only for a workflow input or output, or a value several tasks share.

incomingRefs cache — what PUT does and doesn’t fix:

Scenario Result Fix
New tasks added via PUT incomingRefs generated — task-to-task $var refs resolve immediately None needed
Existing task field values changed via PUT incomingRefs NOT regenerated — literals/changed taskRefs resolve to null Open in Studio → Save
POST /workflow_builder/workflows/save Does NOT regenerate incomingRefs either Open in Studio → Save
Evaluation silently returns false after PUT Stale operand cache Constant-holder workaround below, or Studio save
Workflow hangs at workflow_start (status: running forever) after PUT Any task’s incomingRefs stale Recreate via fresh POST — more PUTs won’t fix it

When you build a workflow document from scratch for POST /workflow_builder/workflows/save, set "scenarios": [], "errors": [], "warnings": [] and "canvasVersion": 3 explicitly. Without canvasVersion, POST /workflow_engine/workflows/validate returns isValid: false with “Task has nodeLocation but canvasVersion is not 3” on every task; both save endpoints still accept the workflow and store those warnings on it, and an import succeeds but stores it as canvasVersion: 1 (the old canvas). A job still starts without any of the four fields (tested on a lab platform), so if jobs/start crashes with "Cannot convert undefined or null to object", look elsewhere first — the adapter list/search tasks note (every declared query parameter must be present in incoming). To create a new workflow, prefer POST /automation-studio/automations, whose documented example body already includes canvasVersion; use workflow_builder/workflows/save only to update an existing, complete document.

Constant-holder workaround (API-only, no Studio save needed): store operand_2 literal values in a newVariable task and reference via {"task": "ca5c", "variable": "value"} — taskRef resolution bypasses the cache.

makeData static input strings do NOT resolve after API create/PUT. The input and outputType fields are backed by job_data (type static). Workaround: use newVariable with value: [...] (array literal) — newVariable.value resolves correctly after API create without a Studio save.

task: "static" values broadly are backed by job_data written at Studio-save time. Any static value (template strings, query paths, model IDs, inline constants in childJob variables dicts) resolves as null at runtime on a freshly API-imported workflow until saved through Automation Studio.

A downstream task can read an upstream task’s output without a job variable. $var.<taskId>.<outVar> resolves whether the upstream task’s outgoing is "", null or a job variable (tested on a live platform). Write outgoing to $var.job.x only when the value is a workflow output or shared by several tasks.

POST /workflow_engine/workflows/validate (6.5.2+) — deep validation before create or update; body {"asset": {...}}, returns {isValid, errors: [], warnings: []}. isValid is errors.length === 0 — warnings don’t count against it, so always check both. See the pre-flight section under “Updating Assets” below for exactly what this endpoint does and doesn’t catch. Use this instead of the older POST /automation-studio/workflows/validate ({"workflow": {...}} → {errors, warnings}, no isValid), which misses errors this one catches.

Nested $var references do not resolve, except for one narrow, named exception list. The engine only substitutes a reference when an incoming value is entirely a $var... string — it does not walk into nested objects/arrays looking for one to substitute (Key Rule 8 in AGENTS.md). The deep validator (above) warns when it finds one. Exceptions — these methods resolve a reference sitting directly on one specific nested key’s values (nothing deeper than one level):

App Method Nested key that’s resolved
GatewayManager runService params
GatewayManager runServiceStatic params
GatewayManager runCode data
AgentSessionManager runAgent inputs
WorkFlowEngine transformation variableMap

Everywhere else — including plain adapter body fields — a $var inside an object or array is passed through as a literal string and never substituted; build the object with merge/makeData/query first instead.

Pre-flight check for the exception table: before saving, flag any runCode.data, runAgent.inputs or transformation.variableMap whose value is a string (a whole-object $var... ref, with or without a #/path suffix) and rewrite it per key; flag a runService.params that doesn’t match its service: an object of $var strings for a service with no declared schema (use a ref to a merge output), or a string for a service that has one (use one $var per key). Also avoid feeding one consumer from both a success-branch merge and an error-branch merge while it reads only one of them: on the other path that reference points at a task that never ran (reasoned from the wiring, not run).


Utility Tasks (WorkFlowEngine)

These are built-in tasks that require no adapter. They handle data manipulation and control flow.

Default for any per-item transform, lookup, or branching loop: runCode + Enable Query below — not a chain of the individual utility tasks that follow. The individual tasks (query, merge, evaluation, forEach, newVariable, makeData, push/pop/shift, deepmerge, transformation, decision, and the rest) are secondary: use one of them alone for a single non-looping operation, or when the per-item work must call another platform task (adapter, app method, childJob) per item — runCode can’t do that, it only runs Python against the data it’s given.

Reshaping data: runCode, not a JST transformation. Put the logic in a runCode task on the canvas — readable Python, testable on its own with JSON on stdin — and have it return everything downstream needs in one result; consumers read its fields with Enable Query ($var.<task>.result#/stdout_json/<field>). The cost: runCode runs on a Gateway5 cluster, so the workflow needs a clusterId input; a transformation runs on the platform. Worked examples: “Process Push Configuration Data” in the push-config assets and the “Standard Output” tasks in vendor-arista-eos.json’s Command Template Runner_v2.

runCode (GatewayManager) — real Python instead of chaining WorkFlowEngine utility tasks

Before wiring a runCode task, read this section and assets/helpers/assets/runcode-taskquery-reference.json in full. Do not construct the task shape from a live job’s error trace or from memory — runCode is a GatewayManager automatic task (not a WorkFlowEngine operation task, an easy but costly mix-up), and its exact field names (clusterId, language, code, data, safety.timeout, packages) are documented in the table below. Getting this wrong produces confusing “additional properties”/“required property” validation errors that look like a project-import problem when the real cause is simply the wrong app/type on this one task.

runCode ships arbitrary Python to a Gateway5 cluster for execution — no pre-configured IAG service needed, unlike runService. It is the single biggest lever for collapsing a forEach + query×N + evaluation + merge×N + push/objectToString/join chain (the kind of per-item transform loop that’s the source of most WorkFlowEngine utility-task gotchas in this file) into 1-2 tasks.

Prerequisites (per official docs):

  • Install Gateway Manager 1.0.10 or later with Platform 6.4.0 or later.
  • Register an active Gateway 5.4.0 or later cluster with Itential Platform using Gateway Manager.
  • Verify that the Python version on the Gateway 5 host is compatible with your scripts and packages. The task uses the default Python interpreter on the host, which is the version invoked by python or python3 on the iagctl command line. Gateway does not enforce a minimum version — run python --version on the Gateway 5 host to check.
  • Confirm that the gateway:code role from GatewayManager is assigned to your group.

When to reach for it: any per-item data transform/lookup/branching loop over a list — the exact shape that otherwise needs forEach (with its job_id-omission and empty-last-transition rules), query per field, evaluation for branching, merge for reassembly, and (for arrays of objects) the objectToString/push/join dance. One runCode task does the whole loop in real Python in a single execution, then one query pulls the result back into a job variable.

Real-world example: a NetBox-devices-to-inventory-nodes mapping (per-device manufacturer lookup, per-vendor secret-path branching, record reshaping) that originally took forEach + 5 query + 1 evaluation + 4 newVariable + 2 merge + push/objectToString/join/makeData (17 tasks) collapsed to 2 tasks (runCode + query), then to 1 task once the trailing query was replaced by an Enable Query decorator too (see below) — with identical, verified output at every stage. Full before/after: assets/helpers/assets/netbox-inventory-sync-runcode-enablequery.json.

Read the real, verified example before building any transform loop:

# See the whole pattern: NetBox devices -> per-vendor secret mapping -> reshaped records,
# via one runCode task + Enable Query, both the "reuse a script" and "native tasks" versions
jq '[.components[] | select(.type=="workflow")] | .[].document.name' \
  assets/helpers/assets/netbox-inventory-sync-runcode-enablequery.json

# Extract the runCode task itself (clusterId, language, code, data, safety, packages)
jq '[.components[].document.tasks // {} | to_entries[] | select(.value.name == "runCode")] | .[0].value' \
  assets/helpers/assets/netbox-inventory-sync-runcode-enablequery.json

# See an Enable Query decorator in place on a real task (top-level AND nested-field examples)
jq '[.components[].document.tasks // {} | to_entries[] | select(.value.variables.decorators // [] | length > 0)] | .[] | {task: .value.name, incoming: .value.variables.incoming, decorators: .value.variables.decorators}' \
  assets/helpers/assets/netbox-inventory-sync-runcode-enablequery.json

Fields (confirmed via live multipleTaskDetails?dereferenceSchemas=true — {"location":"Application","pckg":"GatewayManager","method":"runCode"}):

Field Required Notes
clusterId yes Gateway5 cluster ID to execute on
language yes Enum — currently only "python"
code yes The Python source as a plain string
data yes Object passed to the script as JSON on stdin. Author it per key — {"k": "$var.<task>.<out>", ...} — not as a whole-object string ref to a merge output, which Studio can’t edit (see $var Resolution Rules). Reference the producing task directly (e.g. $var.b1b3.response / $var.b1b3.error) so no merge is needed
safety.timeout yes Seconds, 1-1000
packages no Array of pip requirement strings (e.g. "requests==2.31.0"), installed before execution — omit if stdlib-only

Size limit (per official docs): code is limited to 2 MB. You cannot save code that exceeds this limit. The limit applies to code text only — installed packages do not count toward it.

Package caching (per official docs): the first time you run a task with a new set of packages, the runner installs them into a virtual environment. Subsequent runs reuse the cached environment, so there’s no install overhead. Any change to the packages list triggers a fresh install.

Script contract (the “IAG runCode standard” — see real examples in assets/helpers/assets/vendor-cisco-ios.json and assets/helpers/assets/vendor-juniper-junos.json):

import sys, json
data = json.load(sys.stdin)          # read the `data` object
# ... do the work ...
print(json.dumps(result))            # last line to stdout is the result

Anything printed to stderr is captured separately and does NOT pollute the parsed result — use it for debug logging inside the script if needed.

This exact boilerplate (import sys, json / data = json.load(sys.stdin)) is required in every runCode script — the incoming data field is NOT injected as a pre-parsed variable, it must be read from stdin. A script that references data.get(...) without first loading it from stdin fails at runtime with NameError: name 'data' is not defined. If you’ve already written one correct runCode script in this session, copy its stdin-reading boilerplate verbatim into every subsequent runCode task rather than re-deriving the input-handling logic from assumptions each time — a model that gets this right once and then writes a different, incorrect version for the next runCode task has not actually learned the contract.

Outgoing result object: stdout (raw string — use when the consumer needs a JSON string, e.g. an IAG service’s string-typed param), stdout_json (already-parsed JSON — use when the consumer wants a real array/object, e.g. a task field typed array/object; absent/null if stdout wasn’t valid JSON), stderr, return_code (0 = success), status ("completed"/"error"), started_at/finished_at/elapsed_ms. Pick stdout vs stdout_json based on what the NEXT task’s field type actually wants — don’t always reach for stdout_json by default.

Errors and timeout (per official docs): the task catches unhandled Python exceptions and captures the traceback in result.stderr — the task still completes and the workflow does NOT follow the error transition for an in-script exception. safety.timeout stops execution if exceeded; when exceeded, result.status becomes "error" and result.error contains the platform-specific failure reason. The timeout applies to your code only, not to package installation. If the task cannot run on the gateway at all (not enabled, lost connectivity), the task itself fails and no result output is produced — that error is visible in the task’s Error tab in Operations Manager instead, as a distinct {state, domain, code, message} shape (e.g. "message": "Gateway with id test-code-task is not enabled").

This is exactly why a job showing workflow_end: complete is not proof the workflow worked (see assets/AGENTS.md Rule 28) — an in-script exception in runCode does NOT trigger the task’s error transition, so the workflow sails through to a “successful” completion while result.status is "error" and result.stderr holds a traceback. Always check result.status/result.stderr on every runCode task in a job before reporting the run as successful.

Execution status vs. exit code (per official docs) — three genuinely different failure modes, don’t conflate them:

  • status: "completed" + return_code: 0 — the script ran and exited cleanly.
  • status: "completed" + a non-zero return_code — the script ran but exited with a failure code; your code executed, the script itself signaled failure.
  • status: "error" — a platform-level failure (e.g. a timeout). The script may not have run, or it stopped mid-execution; check result.error. This is distinct from a Gateway connectivity failure (see above), where the task itself errors and no result fields are populated at all.

stdout_json absent after a successful run (per official docs): if result.stdout_json doesn’t appear after return_code: 0, the most common causes are (a) stdout contained non-JSON content, a mix of JSON and other output, or no output, or (b) a value in the output object wasn’t JSON-serializable, so json.dumps() raised a TypeError and wrote a traceback to stderr instead of JSON to stdout. Inspect result.stdout and result.stderr directly to tell which applies.

Gotcha: secret-vault placeholder strings ($SECRET_.../$KEY_...) written as literal Python string constants inside code are NOT resolved by runCode or by the workflow engine — they pass through as plain text, identical to writing the same literal into a newVariable task. That’s expected; the vault resolves them later, at actual device-connection time, not during this task’s execution. This is not the only option, though — per official docs, runCode has its own external-secret mechanism: reference $GATEWAYSECRET_(alias-name) as a literal string in your script, and Gateway resolves the alias at execution time. The secret value is never sent back to Itential Platform. This requires a secret provider/alias already set up on the Gateway side — not yet confirmed live in this repo, but documented here since it directly answers “is there any safe way to use secrets in a runCode script.”

“Enable Query” — inline query decorator on any incoming field (aka “in-task-query”, called “task query” in official docs)

A per-field Studio toggle, not a task or endpoint — invisible to tasks.json/openapi.json/multipleTaskDetails searches. Appears under any incoming field with Variable Source: Job or Task selected (a checkbox labeled “Enable Query”). Lets you query a nested path out of that field’s resolved value inline, eliminating a separate query task that would otherwise sit between the producing task and the field that needs one piece of its output.

Confirmed shape (via a live before/after diff of a Studio edit — not guessed):

  "incoming": {
-   "someField": "$var.job.someVar",
+   "someField": "$var.job.someVar#/nested/path",
    ...
  },
- "decorators": []
+ "decorators": [
+   { "type": "query", "pointer": "/incoming/someField", "displayPath": ".nested.path" }
+ ]
  • The field’s existing $var.job.<name> or $var.<taskId>.<name> reference gets a #/<path> suffix appended directly onto the string (URL-fragment style, /-delimited).
  • The task’s decorators array (always present, usually []) gets one entry: pointer is JSON-Pointer-style pointing at the field (/incoming/<fieldName>); displayPath is the dot-notation the Studio UI shows — its .-separated segments map 1:1 to the #/... suffix’s /-separated segments (.result.stdout → #/result/stdout).
  • Doesn’t create a task, doesn’t touch outgoing — purely a decorator on an existing field’s existing value.

Verified working at runtime, not just in Studio: deleted a query task whose only job was extracting one field from another task’s object output, and instead pointed the downstream consumer’s own field at the source directly with the #/... suffix + decorator. The job variable the deleted task used to populate was confirmed absent from the job’s final variable dump (proving the task really was gone), yet the consuming task still received the correct nested value and the job completed successfully with real data flowing through.

Confirmed scope: works on both a top-level plain-string incoming field AND a field nested inside an object (e.g. runCode’s data.devices) — confirmed by deleting the upstream query task entirely and verifying the job variable it used to write was absent from the job’s final state while the downstream task still received the correct value. For a nested field, the #/<path> suffix goes on the nested value itself ("data": {"devices": "$var.job.someVar#/nested/path"}) and pointer uses the full nested path (/incoming/data/devices).

Where pointer goes — verified against Studio saves, per field shape

Everything in this subsection was verified on Itential Platform 6.5.2 (Automation Studio 4.69.69), 2026-10: Studio saves for the stored shapes, live jobs for the runtime behavior. Re-check on other versions before relying on an edge case.

The rule is the same everywhere: pointer is /incoming/ plus the path to the field that holds the reference string; the value gets #/<path>; displayPath is the same path with dots (#/body/name → .body.name). The holder moves for nested shapes, so check the table instead of assuming /incoming/<field>:

Field shape Examples Value stored pointer
Top-level field (string, array, object or “any” typed) toLowerCase.str, newVariable.value, stub.response, join.arr, forEach.data_array, setObjectKey.obj, ViewData.body, restCall.uri/body, ShowJsonForm.instance_data, eventListenerJob.topic/schema, adapter reference $var.job.x#/body/name /incoming/<field>
A whole object field makeData / ViewData / ViewHTML / renderJinja2TemplateWithCast variables; adapter requestBodyPayload $var.job.x#/body /incoming/variables
One key inside runCode data (the documented one-level exception) runCode.data.x $var.job.x#/body /incoming/data/x
One input inside transformation variableMap (also one level deep) variableMap.<input> $var.job.x#/body/list /incoming/variableMap/<input>
merge item data_to_merge[i].value {"task":"job","variable":"x#/body/name"} /incoming/data_to_merge/<i>/value/variable
childJob variable variables.<name> {"task":"job","value":"x#/body/name"} /incoming/variables/<name>
childJob loop input (loopType set) data_array $var.job.x#/body/devices /incoming/data_array
evaluation operand operand_1, operand_2 no decorator — use query / rightQuery on the evaluation object —

Array indexes: the value uses the index as a path segment and displayPath uses brackets — $var.job.x#/body/items/0/name with displayPath .body.items[0].name (whole element: #/body/items/1 → .body.items[1]). Variadic inputs (stringConcat stringN, arrayPush elementN, assign sourceN) are stored under that literal key, so the pointer is /incoming/stringN etc. A merge item reading a childJob output keeps "value": "job_details" and gets the query on an added variable key — {"task": "<childJobId>", "value": "job_details", "variable": "job_details#/received"}, pointer /incoming/data_to_merge/<i>/value/variable. Putting the #/path on value instead resolves to null with no error.

Keys with special characters: / and ~ in a key use JSON Pointer escapes in the value (#/a~1b, #/a~0b) and appear unescaped in displayPath (.a/b, .a~b). A key containing . uses bracket form: the value keeps the dot inside one segment (#/a.b, nested #/body/mgmt.ip) and displayPath brackets that key — .["a.b"] at the top level, .body["mgmt.ip"] when nested (no dot before the bracket) — which is what Studio writes when you enter those paths; it resolves the dotted key. Other forms are accepted but produce a different path: ."a.b" → #/"a/b" and .a\.b → #/a\/b both fail at runtime, and .a.b → #/a/b reads the nested a → b with no warning. If you build one by API, use the bracket displayPath so Studio shows it correctly.

A reference to another task’s output uses the same forms with the task id in place of job — {"task":"a001","variable":"value#/body/ip"} for a merge item, {"task":"a001","value":"value#/body/ip"} for a childJob variable, $var.a001.value#/body/ip for a plain or runCode field — with the same pointer. The path goes after the output name (value#/body/ip).

Note the merge and childJob references carry no $var. prefix (x#/body/name, same as their plain refs), and the childJob pointer stops at the variable name while the merge pointer includes /value/variable.

Both halves are required, and they are read by different things. The job runs from the #/path suffix; Studio displays from the decorators entry. Writing only the suffix (or a decorator whose pointer is one level off) still runs correctly, so it passes every job test — but Studio shows the field with no query text, and opening then saving the workflow silently strips the #/path. The task then receives the whole object instead of the extracted value and the job still completes with no error (a downstream query or evaluation on it is what fails). A correct decorator survives saves unchanged.

A path that doesn’t exist at runtime fails loudly — except on evaluation. On plain fields, merge items, childJob variables (the child job is never started) and runCode data keys, the task takes its error transition with "Query failed for task incoming variable: <source variable>." — wire an error transition. On evaluation, a missing query path takes the failure transition, which looks exactly like a comparison that was false; check the path when an evaluation unexpectedly fails.

Do not guess a merge or childJob form. Adding a query key beside variable/value (the way an evaluation operand looks) is ignored at runtime and invisible in Studio — the whole object is passed through with no error. Use the #/path forms in the table.

Build objects upstream; never put a $var inside a static object to give a field a query. Reference the whole field (variables, requestBodyPayload) and let the query shape it, or build the object with merge/makeData first. runCode data is the one place a $var resolves one key deep.

Generate and verify decorators with the helper instead of writing them by hand: python3 assets/helpers/enable_query.py check workflow.json lists every task whose decorators are missing, dangling or off by a level, and flags the ignored query-key forms above and any $var placed inside a static object (which is sent as literal text); fix workflow.json rewrites the query decorators from the references, keeps any other decorators (e.g. encryption), and lists what it can’t repair — those need a change to the reference itself. fix writes the file back as 2-space JSON, so on a hand-formatted file such as one in assets/helpers/assets/, apply check’s findings by hand. In code: apply_decorators(task). Run check before every create/update.

When to reach for it: any spot where a query task’s only purpose is pulling one field out of an object and handing it to exactly one downstream task’s field. Skip it if the queried value feeds more than one consumer, or needs further transformation (evaluation, string ops) before use — a real query task is still the right call there.

Validator vs runtime: on some platform builds, POST /workflow_engine/workflows/validate reports a query decorator as a schema error (decorators/0/type: must be equal to one of the allowed values — that build’s schema only allows encryption) and warns that the field gets an object. Jobs still resolve the decorated reference at runtime — verified by a sendConfig whose config and inventory came through decorators and pushed the right line to a live device. Don’t “fix” those errors by removing the decorators.

evaluation does not use a decorator. Its operands are structured references ({"task":"job","variable":"x"}), not $var strings, so there is no #/path suffix and no decorators entry. The query lives on the evaluation object itself: query (applied to operand_1) and rightQuery, exactly as Studio writes them — see ### evaluation below. Do not put a query key inside the operand object: it is ignored at runtime and Studio shows nothing for it.

Worked example: wiring runCode + task query together

A complete walkthrough of bringing multiple upstream tasks’ data into a runCode script, using task query to unwrap each one, and reading the result back out — confirmed live end-to-end. Real example (two NetBox list calls combined into one summary): assets/helpers/assets/runcode-taskquery-reference.json.

1. Bring variables into runCode via its data field. Each key in incoming.data becomes one entry your script receives on stdin. The value is a reference expression, same syntax as any other task field:

"data": {
  "devicelist": "$var.8850.response",
  "interfacelist": "$var.4e04.response"
}

This hands the entire response output of tasks 8850 and 4e04 to the script, under the keys devicelist/interfacelist. No query needed yet — this alone is enough if the script itself can navigate whatever shape arrives (see step 4).

2. Design a task query only if you want to unwrap a nested field before the script sees it. Append #/<json-pointer-path> to the reference string, and add a matching entry to the task’s decorators array — both parts are required and must agree:

"data": {
  "devicelist": "$var.8850.response#/body"
},
"decorators": [
  { "type": "query", "pointer": "/incoming/data/devicelist", "displayPath": ".body" }
]
  • The #/<path> suffix goes on the value string itself, /-delimited (JSON-Pointer style).
  • The decorator’s pointer is /incoming/data/<key> (locates which incoming field has a query), and displayPath is the Studio-UI-facing dot-path (.-delimited, 1:1 with the #/... segments).
  • This query is evaluated once, before the script ever runs. If the path doesn’t exist on the real resolved object, the task takes its error transition before the script runs, with "Query failed for task incoming variable: <source variable>." (the job stops there if the task has no error transition). The script never gets a chance to handle a missing/wrong path defensively; get the path right up front.

Don’t guess the path — verify the real shape first. The platform’s own job-detail API does not help here: GET /operations-manager/jobs/{id} always returns the static workflow definition (empty-string outgoing placeholders) regardless of include/dereference params or job completion status — confirmed directly, including via the browser UI’s own network call using session-cookie auth, not just a scripted API token. The two ways that do work: (a) open the job in Operations Manager / Studio and read the task’s real Output/Response value directly in the UI, or (b) temporarily pass the reference through without a query, and have the script itself print the shape it received (type(), keys, a sample item) so you can see the real structure before committing to a query path.

Confirmed real-world case: a REST/HTTP-style adapter task whose outgoing field is named response returns the full HTTP response object, not the API payload directly — the payload is at response.body (response itself is {body, headers, statusCode, ...}-shaped). This is a different wrapping than the existing documented result.response convention elsewhere in this file (different outgoing field name, different nesting depth) — don’t conflate the two. Confirmed on NetBox’s dcim_devices_list/dcim_interfaces_list: querying #/results directly failed (that path doesn’t exist at the top level); the real path is #/body, landing on {count, next, previous, results: [...]}}.

The query must be the field’s entire value — it does not work embedded inside a larger static object. Confirmed by feeding runCode’s output into a second, non-runCode task (extras_tags_create) requiring a requestBodyPayload object with name/slug/description:

  • Failed silently: requestBodyPayload set to a static object literal with the query on one of its own keys — {"name": "my-tag", "slug": "my-tag", "description": "$var.3c0b.result#/stdout_json/devices/0/name"} + a decorator pointing at /incoming/requestBodyPayload/description. The query never resolved. Worse, this didn’t fail loudly: the literal, unresolved $var...#/... string was sent through as-is, and NetBox — whose description field accepts any free text — happily stored the garbage literal string as if it were real data. The failure only became visible when the same pattern was tried on a strictly-typed field (weight, an integer): NetBox rejected the literal string with "A valid integer is required", which looked like a type/casting problem but was actually a masked query-resolution failure. A task query that “succeeds” into a loosely-typed field is not proof the query itself worked — check a strict field, or inspect the literal value sent, before trusting it.
  • Confirmed working: shape the upstream task’s output to already match the downstream field’s whole expected structure, then query the entire field in one shot — "requestBodyPayload": "$var.3c0b.result#/stdout_json/tags_payload" (where tags_payload in runCode’s own JSON output is already {name, slug, description}-shaped), decorator {"pointer": "/incoming/requestBodyPayload", "displayPath": ".stdout_json.tags_payload"}. Verified end-to-end: real HTTP POST to NetBox, real object received (confirmed via a genuine "tag with this name already exists" conflict on a repeat run — proof the resolved payload’s name/slug were real values, not placeholders).
  • This also confirms task query is not runCode-specific — it works the same way on a plain adapter task, as long as the whole-field rule above is followed.

3. Access the variables inside the script. Everything in data arrives as one JSON object on stdin, keyed exactly as configured:

import sys, json
data = json.load(sys.stdin)
devicelist = data['devicelist']    # already whatever step 2's query resolved it to, if a query was used
interfacelist = data['interfacelist']

If a task query already unwrapped a field (e.g., to .body), the script receives that unwrapped value directly — it doesn’t need to re-navigate past a level the query already resolved. If no query was used, the script receives the full raw reference and is responsible for navigating it itself (defensive code that handles more than one possible shape is reasonable here, since a script CAN branch/fall back at runtime, unlike a query decorator which fails hard on a wrong path).

4. Return the result. The last line printed to stdout must be json.dumps(...) of whatever the next task (or the workflow’s output) should consume — this becomes result.stdout_json. Anything else printed for debugging goes to stderr instead, so it doesn’t corrupt the parsed result (see the runCode fields table above).

Individual utility tasks (secondary — single operations, or when runCode doesn’t apply)

query

Extract nested values from objects using dot-path syntax.

Incoming: pass_on_null (boolean), query (string — dot-path), obj (object — usually $var ref) Outgoing: return_data (any) Transitions: success (found), failure (null/undefined when pass_on_null: false)

{
  "incoming": {
    "pass_on_null": false,
    "query": "response.id",
    "obj": "$var.a1b2.result"
  },
  "outgoing": {
    "return_data": "$var.job.changeId"
  }
}

IMPORTANT: Don’t guess the query path for adapter responses. Adapters transform upstream API responses — the field path in the adapter’s output is NOT the same as the native API’s response structure. The adapter’s result outgoing is always a {response, headers, metrics} object, never a primitive. When the upstream API returns a simple string (like Infoblox’s _ref), it’s at result.response, not result directly. Always verify the actual response shape from a test job (GET /operations-manager/jobs/{jobId} → data.tasks) before wiring a path.

pass_on_null: false does not reliably trigger the failure transition for an out-of-bounds array index. Live-tested: query: "response.results[0].status.value" against an adapter response where results resolved to an empty array ([] — e.g. a NetBox lookup for a device name with zero matches) produced return_data: null but the task still took the success transition, not failure. This is different from a missing/undefined key partway through the path, which does trigger failure as documented. If a downstream task (e.g. evaluation) needs to distinguish “found but null” from “not found” for an array-index path, don’t rely on the query task’s own success/failure transition — check the extracted value explicitly (e.g. an evaluation testing the job variable for null/empty) rather than assuming a failure transition will catch it.

merge

Build an object from multiple resolved values. Primary workaround for $var not resolving inside nested objects.

Incoming: data_to_merge (array, min 2 items) Outgoing: merged_object (object)

IMPORTANT: The field is "variable" NOT "value" in the reference objects inside data_to_merge.

Reference format in data_to_merge:

  • {"task": "job", "variable": "varName"} — pull from a user-supplied job variable (input to the workflow)
  • {"task": "static", "variable": "literalValue"} — literal value
  • {"task": "taskId", "variable": "outVar"} — pull from a previous task’s output

WARNING — {task:"job"} references add fields to inputSchema.required. The platform scans every data_to_merge entry in merge tasks (and every variables entry in childJob) for {task:"job"} references and automatically adds that variable name to inputSchema.required. This means using {task:"job", variable:"changeId"} for a variable that was produced internally by a query task will prompt operators to supply changeId as a workflow input — even though it should never come from the user.

Rule: only use {task:"job"} for variables that are genuine workflow inputs. For anything produced by an earlier task, use the producing task’s ref directly:

Value source Correct ref form
User workflow input {"task": "job", "variable": "x"}
query output {"task": "queryTaskId", "variable": "return_data"}
merge output {"task": "mergeTaskId", "variable": "merged_object"}
newVariable output {"task": "newVarTaskId", "variable": "value"}
makeData output {"task": "makeDataTaskId", "variable": "output"}
parse output {"task": "parseTaskId", "variable": "return_data"}
adapter task output {"task": "adapterTaskId", "variable": "result"}
{
  "incoming": {
    "data_to_merge": [
      {"key": "hostname", "value": {"task": "static", "variable": "IOS-CAT8KV-1"}},
      {"key": "details", "value": {"task": "job", "variable": "deviceInfo"}},
      {"key": "config", "value": {"task": "a1b2", "variable": "renderedTemplate"}}
    ]
  },
  "outgoing": {
    "merged_object": "$var.job.requestBody"
  }
}

Gotchas: Requires at least 2 items (1 item = silently null). Outgoing MUST declare "merged_object": null (empty {} makes it unreachable). Duplicate keys produce arrays — merging {"ip": "1.2.3.4"} and {"ip": "1.2.3.4"} yields {"ip": ["1.2.3.4", "1.2.3.4"]}, not an overwrite. To avoid this, pass a pre-built object as a single workflow input variable instead of merging multiple objects with the same keys.

parse

Convert a JSON string into a JavaScript object. Essential after extracting result.stdout from runService (which is always a string, even when the script printed valid JSON).

Incoming: stringToParse (string — the JSON string to parse) Outgoing: result (object — the parsed object)

{
  "name": "parse",
  "canvasName": "parse",
  "summary": "Parse JSON String",
  "location": "Application",
  "locationType": null,
  "app": "WorkFlowEngine",
  "type": "operation",
  "displayName": "WorkFlowEngine",
  "variables": {
    "incoming": {
      "stringToParse": "$var.a1b2.return_data"
    },
    "outgoing": {
      "result": "$var.job.parsedOutput"
    }
  },
  "actor": "Pronghorn"
}

Common pattern — runService → query → parse:

runService → query(result.stdout) → parse(stringToParse) → use parsed fields

After parse, fields are accessible: $var.parseTask.result.hostname, $var.parseTask.result.status, etc.

evaluation

Conditional branching. MUST have BOTH success AND failure transitions.

Incoming: all_true_flag (boolean), evaluation_groups (array) Outgoing: return_value (boolean) Transitions: success (true), failure (false)

Operator enum — closed set. Only these 8 are valid:

contains, !contains, <, <=, >, >=, ==, !=

regex, match, matches, contains_key, in, startsWith — do not exist. An invalid operator silently returns false with empty outgoing and finish_state: failure. No error message. Always validate against this list before wiring. Source of truth: openapi.json at components/schemas/workflow_engine_wfEngineCommon_evaluationItem/properties/operator/enum.

contains is regex-based, not substring. operand_2 is interpreted as a regex pattern. A literal like 9.2(4) is parsed as regex — . matches any char, (4) becomes a capture group — and may match unintended strings or fail to match the intended one. Escape regex metacharacters in literal patterns: 9\.2\(4\) not 9.2(4).

contains also works for object-key presence — it is the universal “does X contain Y” operator. On a string operand it does regex matching; on an object operand it tests key presence. There is no separate contains_key operator.

Direct evaluation test (no workflow needed):

POST /workflow_engine/runEvaluationGroups
{"evaluation_groups":[{"operator":"AND","evaluations":[{"operand_1":"<test input>","operator":"contains","operand_2":"<pattern>"}]}]}

Returns true/false. Invalid operators silently return false. Use this to validate operators and escape patterns before wiring them into a workflow.

incomingRefs cache — API PUT does not regenerate it for existing task changes. See the incomingRefs table in $var Resolution Rules. Diagnostic sign: GET /operations-manager/tasks/{iterationUUID} shows incomingRefs[n].taskId: null or taskPointer: "/variables/outgoing/undefined".

Operand reference format (uses "variable", same as merge):

  • {"task": "job", "variable": "varName"}
  • {"task": "static", "variable": "literalValue"}
{
  "incoming": {
    "all_true_flag": true,
    "evaluation_groups": [{
      "all_true_flag": true,
      "evaluations": [{
        "operand_1": {"variable": "status", "task": "job"},
        "operator": "==",
        "operand_2": {"variable": "success", "task": "static"}
      }]
    }]
  },
  "outgoing": {"return_value": null}
}

Drill into a nested field with query on the evaluation object — no separate query task needed. Put query (applies to operand_1) and rightQuery (applies to operand_2) next to operator on every evaluation item — each item in each group carries its own pair, in any position, using the dot-path syntax Studio writes (.body.name):

{
  "query": ".body.name",
  "operand_1": {"task": "job", "variable": "src"},
  "operator": "==",
  "operand_2": {"task": "static", "variable": "poc-a"},
  "rightQuery": ""
}

Always include rightQuery (empty string when unused), as Studio does. At runtime the leading dot is optional and single- or multi-segment paths both work; write the Studio form so a Studio save changes nothing. This is the form Studio shows under Enable Query, and it survives a save. No decorator is involved.

Do not put query inside the operand ("operand_1": {"task": "job", "variable": "src", "query": "body.name"}). Tested with the operand pointing at a job variable and at a task output, single and multi-segment, with and without a leading dot: the query was ignored every time — operand_1 resolved to the whole object, so the comparison silently failed — and Studio shows no Enable Query for it. If you inherit a workflow that uses it, move the query to the evaluation object. If a query still has no effect, fall back to a standalone query task writing a job variable, then compare that.

childJob

Run another workflow as a sub-job. Read a live childJob example first:

jq '[.components[].document.tasks // {} | to_entries[] | select(.value.name == "childJob")] | first | .value' \
  assets/helpers/assets/vendor-cisco-ios.json

Critical differences from normal tasks:

  • actor MUST be "job" — not "Pronghorn"
  • task MUST be "" (empty string)
  • outgoing.job_details MUST be null — do NOT override with $var.job.X
  • All incoming fields required — even unused ones: "data_array": "", "transformation": "", "loopType": ""

childJob cannot resolve project-scoped workflow names. {"workflow": "@<projectId>: <name>"} fails at runtime with "Error starting child job: Cannot find workflow ..." — and so does the bare unprefixed name ({"workflow": "<name>"}), even when the calling workflow is moved into the same project as the target. Confirmed with an isolated minimal test: a standalone parent calling a standalone (unprojected) child by bare name succeeds (fails later for an unrelated reason, but the child IS found); the identical parent calling a project-scoped child by either name form fails to find it. Ruled out incomingRefs staleness as the cause — persisted after a full delete-and-recreate of the parent. If you need to call a workflow that lives in a project, inline its task(s) directly into the calling workflow instead of childJob-ing it — there is no known workaround for calling it as a child.

A merge item reading a childJob’s output uses "variable" — {"task": "<childJobId>", "variable": "job_details"}, the same key as any other merge item. With "value" instead, the workflow creates, updates and runs without error but the item resolves to null. (When you enable a query on this item in Studio it keeps a "value": "job_details" key and adds "variable": "job_details#/<path>"; the runtime reads variable.) Tested on Itential Platform 6.5.2.

Variables use {"task", "value"} syntax — NOT $var:

{
  "incoming": {
    "task": "",
    "workflow": "My Child Workflow",
    "variables": {
      "deviceName": {"task": "job", "value": "deviceName"},
      "configData": {"task": "a1b2", "value": "return_data"}
    },
    "data_array": "",
    "transformation": "",
    "loopType": ""
  },
  "outgoing": {"job_details": null}
}

childJob uses "value", not "variable" — see Variable Syntax Reference below for the full merge/evaluation contrast.

Variable passing:

  • {"task": "static", "value": [...]} — literal value
  • {"task": "job", "value": "varName"} — parent job variable (must exist at start)
  • {"task": "taskId", "value": "outVar"} — previous task’s output (preferred for runtime data)

Loop modes: loopType: "" (single), "parallel" (multiple simultaneous), "sequential" (one at a time). With loops, use data_array (each element becomes a child job’s variables) and set variables: {}.

Querying childJob output:

{
  "name": "query",
  "variables": {
    "incoming": {
      "query": "taskStatus",
      "obj": "$var.f48f.job_details",
      "pass_on_null": false
    }
  }
}

Use flat variable names, NOT nested paths. For loop output: "[**].fieldName".

forEach

Iterate over an array. Deprecated for new builds — prefer childJob with loopType for calling another workflow per item, or runCode (see Step 0a above) for a per-item data transform/branch that doesn’t need to call another platform task. Still common in existing workflows, and this section stays detailed because you’ll need to read/modify existing forEach loops even when not authoring new ones.

Incoming: data_array (array) — ONLY data_array. Do NOT include job_id in incoming — it triggers errors. Outgoing: current_item (any)

Transition pattern (critical):

forEach --state:loop--> firstBodyTask -> ... -> lastBodyTask --(empty {})
forEach --state:success--> nextTaskAfterLoop

forEach constraints (all four are required)

  1. incoming must only contain data_array — do NOT include job_id or any other field. Adding job_id causes errors at runtime.

  2. $var.<taskId>.<output> does NOT resolve inside the loop body — string references like $var.a01.current_item silently resolve to null inside a forEach body. Use $var.job.<varName> instead (bind the forEach’s outgoing to a job variable and reference that). This applies to ALL reference styles — even taskRef objects {"task": "outerTask", "variable": "current_item"} are unreliable inside a nested body.

  3. Loop body tasks cannot transition to tasks outside the loop — no error transitions from loop body tasks to external error handlers. The forEach task itself handles exit via state: "error" on the forEach transition. Handle errors within the loop body, then let the forEach’s error transition route out.

  4. The last loop body task signals loop-back with an empty {} transition — do NOT add an explicit loop-back target pointing to forEach.

"transitions": {
  "forEachTaskId": {
    "firstBodyTask": {"type": "standard", "state": "loop"},
    "nextTaskAfterLoop": {"type": "standard", "state": "success"},
    "errorHandlerTask": {"type": "standard", "state": "error"}
  },
  "lastBodyTask": {}
}

newVariable

Create or set a job variable at runtime.

Incoming: name (string), value (any) Outgoing: value (any)

{
  "incoming": {"name": "taskStatus", "value": "success"},
  "outgoing": {"value": "$var.job.taskStatus"}
}

GOTCHA: value resolves only a whole-field reference — "$var.job.x", or with Enable Query "$var.job.x#/path" plus its decorator. A $var embedded in a longer string ("prefix-$var.job.x") or placed inside an object/array value is stored as literal text. Use merge (or a template) to build dynamic values. Tested on Itential Platform 6.5.2.

makeData

Construct data with <!var!> variable substitution.

Incoming: input (string with <!var!> placeholders), outputType ("string"/"json"/"number"/"boolean"), variables (object) Outgoing: output (any)

The variables field must be a resolved object. Use merge first to build it, then pass via $var.taskId.merged_object:

merge (build variables object) → makeData (use $var.taskId.merged_object as variables)

WARNING — makeData.incoming.variables cannot use $var references to a merge that sources childJob output. When a merge task’s data_to_merge contains a childJob reference (e.g., {"task": "childJobId", "variable": "job_details"}), the platform cannot compile $var.<mergeId>.merged_object as a taskRef for makeData.incoming.variables — it is stored as a literal static string. Template substitution then operates on the literal string and emits unresolved placeholders.

query.incoming.obj does NOT have this limitation — it resolves $var.<mergeId>.merged_object correctly even when the merge references childJob output.

Fix: extract individual values from the childJob-sourced merge using query tasks, then pass those resolved scalars to makeData via a second merge (that contains only non-childJob refs). Do NOT feed a childJob-sourced merge directly into makeData’s variables.

delay

Pause execution. Incoming: time (integer, seconds). Outgoing: time_in_milliseconds.

push / pop / shift

Array manipulation on job variables by name (plain string, NOT $var reference).

{
  "incoming": {
    "job_variable": "collectedResults",
    "item_to_push": "$var.c3d4.return_data"
  }
}

GOTCHA: Pass "myArray", NOT "$var.job.myArray".

deepmerge

Same as merge but merges nested objects recursively instead of overwriting top-level keys. Use when combining objects that share nested keys.

Incoming: data_to_merge (array, min 2 items — same format as merge) Outgoing: merged_object (object)

transformation

Perform JSON transformation using JST (JSON Schema Transformation).

Incoming: tr_id (string — transformation ID), variableMap (object — maps transformation inputs to data locations), options (object, optional — e.g., {"extractOutput": true}) Outgoing: outgoing (any)

Used in childJob mode 3 (loop with transformation) to reshape each data_array element before passing to the child.

Create transformations before the workflows that call them, and use the ids the platform returns for tr_id. Transformation ids are not portable between platforms, so a copied or invented tr_id points at nothing. The validator warns (“Transformation tasks must reference an existing transformation”) when the id isn’t in the project.

decision

Multi-way branching based on conditions. Unlike evaluation (binary true/false), decision branches to different tasks based on multiple conditions.

Incoming: decisionArray (array of decision objects with conditions and target task IDs) Outgoing: return_value (string — the ID of the next task)

restCall

Make external HTTP calls from within a workflow. Use when calling APIs not exposed through adapters.

Before reaching for restCall, search tasks.json for a native app task first — jq '.[] | select(.app=="GatewayManager")' tasks.json (or InventoryManager, or the target adapter’s type name). restCall against the Itential platform’s own internal REST API (e.g., /gateway_manager/v1/services/run, /automation-studio/...) is almost always the wrong choice — a native task exists for this (e.g., GatewayManager.sendCommand/runService, InventoryManager.buildInventoryFilter) and handles auth, response shaping, and validation for you. Reach for restCall only for genuinely external third-party APIs that have no adapter and no native task.

Response shape — no wrapper. restCall returns the already-parsed JSON body directly as the outgoing value. There is no response or result wrapper. Query paths target body fields directly:

Correct:   "query": "access_token"
Wrong:     "query": "response.access_token"   ← no response wrapper
Wrong:     "query": "result.access_token"     ← no result wrapper

This is the opposite of adapter tasks (e.g., genericAdapterRequest), which always wrap the upstream response in {response, headers, metrics}. Don’t cross-apply the adapter query paths to restCall output — you’ll get null every time.

modify

Modify data by querying into an object and replacing with a new value.

Incoming: object_to_update (any), query (string — json-query path), new_value (any) Outgoing: updated_object (any)

validateJsonSchema

Validate JSON data against a JSON schema.

Incoming: jsonData (object), schema (object) Outgoing: result (object — {"valid": true} or {"valid": false})

Additional Utility Tasks (60+)

Search tasks.json for the full catalog:

jq '.[] | select(.app == "WorkFlowEngine") | {name, summary}' {use-case}/tasks.json
Category Examples
String stringConcat, replace, split, toLowerCase, toUpperCase, trim, substring
Array arrayConcat, arrayPush, sort, join, arraySlice, map, reverse
Object assign, keys, values, objectHasOwnProperty, setObjectKey
Time getTime, addDuration, convertTimezone, calculateTimeDiff
Parse/Transform parse, transformation, stringify
Tools restCall, csvStringToJson, excelToJson, asciiToBase64

Reach for purpose-built tasks before chaining primitives. Two tasks that are commonly underused:

  • setObjectKey (WorkFlowEngine) — writes a value directly into a nested key of an existing object. Use instead of query + merge when updating a single field on an object already in $var.job.*.
  • renderJinja2ContextWithCast (ConfigurationManager) — renders a Jinja2 template with the full job context automatically injected, plus optional type casting on the output. Use instead of merge → renderJinja2 → query chains when the template needs access to existing job variables. Outputs renderedTemplate accessible via $var.<taskId>.renderedTemplate.

Fetch full schemas with POST /automation-studio/multipleTaskDetails?dereferenceSchemas=true.

Task Endpoint Patterns (Standalone Testing)

Some tasks have standalone REST endpoints — faster than creating test workflows:

  • WorkFlowEngine: POST /workflow_engine/{method} (e.g., /workflow_engine/query) — requires job_id (use dummy ObjectId "4321abcdef694aa79dae47ad")
  • MOP: POST /mop/RunCommandTemplate — test command templates directly
  • TemplateBuilder: POST /template_builder/templates/{name}/renderJinja with {"context": {...}} (note: context, not variables)

Most utility tasks (array ops, string ops, forEach, childJob, merge) do NOT have standalone endpoints. Test those by creating a minimal start → task → end workflow and running via jobs/start.


Templates (Jinja2 / TextFSM)

POST /automation-studio/templates
{
  "template": {
    "name": "VLAN_Interface_Config",
    "type": "jinja2",
    "group": "Cisco IOS",
    "command": "configure terminal",
    "description": "Generates VLAN interface config",
    "template": "interface Vlan{{ vlan_id }}\n description {{ description }}\n ip address {{ ip_address }} {{ subnet_mask }}\n no shutdown",
    "data": "{\"vlan_id\": 100, \"description\": \"Management\", \"ip_address\": \"10.0.1.1\", \"subnet_mask\": \"255.255.255.0\"}"
  }
}

Required fields: name, group, command, description, template, data, type

Give every template field a real value. data must be a non-blank string of sample text to test the template with (for TextFSM, a few lines of the command’s output) — never {}. A missing or non-string data fails that component on import while the project still reports success, so always check metadata.failedComponents after importing (Field 'data' must be a string, Field 'data' is required). A blank command, description, data or body imports but fails the template validator. Validate before importing — see “Validate every asset with its own validator”.

Types: jinja2 (config generation) or textfsm (output parsing)

Test rendering directly:

POST /template_builder/templates/{name}/renderJinja
{"context": {"vlan_id": 100, "description": "Management"}}

Gotchas:

  • group cannot be empty or whitespace-only
  • Use underscores in template names (e.g., IOS_Switchport_Config)
  • data field is a JSON string, not an object
  • Variable syntax is {{ var }} (Jinja2), NOT $var or <!var!>
  • No from_json filter — Ansible’s from_json Jinja2 filter does NOT exist in Itential’s TemplateBuilder. If you need to parse a JSON string, use a parse task before the template render step, not a filter inside the template
  • renderJinjaTemplate as a workflow task — use TemplateBuilder.renderJinjaTemplate with incoming templateName (string) and variables (object). Output is at result.renderedTemplate (string). Different from the standalone API endpoint which uses context instead of variables

Command Templates (MOP)

MOP manages command templates for running CLI commands with validation rules. MOP is read-only validation only — never use it to push config.

To push config to a device, use GatewayManager.sendConfig — not MOP. It sends configuration text to Inventory Manager nodes through a Gateway5 cluster. The standard pattern for any config push delivery is:

Pre-Check (RunCommandTemplate child)
  → Push Configuration to Device (renderJinjaTemplate → approval ViewData → sendConfig → per-node success check)
  → Post-Check (RunCommandTemplate child)
  → runTemplatesDiff (compare pre vs post)

Verified against a live Arista EOS device:

  • config is the rendered text as one string, and inventory is [{"inventory": "<inventory>", "nodeNames": ["<node>", ...]}]. Build both in one runCode task before the push — it takes renderJinjaTemplate’s output (an object, {renderedTemplate: "..."}), the device(s) and the inventory name, and returns {config, inventory, ...}; sendConfig reads them with Enable Query (config: "$var.<task>.result#/stdout_json/config"). See “Process Push Configuration Data” in the asset. (Without runCode: a query task for the text, and merge + arrayPush for the inventory — $var doesn’t resolve inside it.)
  • Config mode needs enable: the node’s itential_driver_options.netmiko.become must be true (plus secret if the device has an enable password). Without it the push fails with ReadTimeout: Pattern not detected: '<host>.*\)\#'.
  • Output: result.result.results[], one {name, host, output, success} per node — there is no overall state field. Check every node: an evaluation with query: "result.results" != [] and query: "result.results[*].success" !contains false (group all_true_flag: true). Prove the check by running it once against a push that can’t land — a wrong query path passes silently.
  • A Gateway4-style template that wraps its lines in conf t … end still works on EOS; the driver enters config mode itself, so new templates don’t need the wrapper.

AGManager.itential_cli / itential_set_config are Gateway4. Don’t build new deliveries on them; use them only while maintaining an existing Gateway4 workflow. When that workflow moves to sendConfig, every task that read the old output (stdout.completed[0].response[*].status, …icode) must change too — two ways:

  1. Re-point each consumer to the new shape (the evaluation above, a query on result.results[0].output for a ViewData body). Best when only a few tasks read it.
  2. Rebuild the old shape with one runCode task right after sendConfig, then point the consumers at it and prefix their queries with stdout_json. — the query logic itself stays as it was. Best when many tasks or transformations parse the Gateway4 shape:
    import sys, json
    data = json.load(sys.stdin)                      # data: {"result": "$var.<sendConfig task>.result"}
    nodes = ((data.get("result") or {}).get("result") or {}).get("results") or []
    print(json.dumps({"completed": [{"icode": "AD.200", "response": [
        {"host": n.get("name"), "status": "SUCCESS" if n.get("success") else "FAILURE",
         "stdout": n.get("output", "")} for n in nodes]}]}))
    runCode’s own result is flat ({status, return_code, stdout, stderr, stdout_json}) — no JSON-RPC envelope, unlike sendConfig’s.

Gateway4’s icode: "AD.200" only means the call reached the gateway — a push the device rejected still returns it. A workflow that checked only icode never caught failed pushes; the per-node check does.

Read the Arista EOS “Push Configuration to Device - IAG” and “Command Template Runner” workflows before building any config push delivery:

jq '[.components[] | select(.type=="workflow") | select(.document.name | test("Push Config|Command Template"))] | .[].document | {name:.name, tasks:.tasks, transitions:.transitions}' \
  assets/helpers/assets/vendor-arista-eos.json

Create a Command Template

POST /mop/createTemplate
{
  "mop": {
    "name": "Port_Turn_Up_Pre_Check",
    "description": "Validates interface and VLAN",
    "os": "",
    "passRule": true,
    "ignoreWarnings": false,
    "commands": [
      {
        "command": "show interface <!interface!>",
        "passRule": true,
        "rules": [
          {
            "rule": "line protocol is",
            "eval": "contains",
            "severity": "error"
          }
        ]
      },
      {
        "command": "show vlan brief",
        "passRule": true,
        "rules": [
          {
            "rule": "<!vlan_id!>",
            "eval": "contains",
            "severity": "error"
          }
        ]
      }
    ]
  }
}

Variable syntax: <!variable_name!> in both commands and rules (NOT {{ }} or $var)

passRule Logic

  • Template-level passRule: true = ALL commands must pass (AND)
  • Template-level passRule: false = ONE command must pass (OR)
  • Command-level = same logic for rules within a command

Rule Evaluation

Eval Purpose Example
contains String exists in output "line protocol is"
!contains String does NOT exist "ERROR"
contains1 String exists exactly once "Active"
RegEx Regex matches (capital R, E!) "/\\d+\\.\\d+/"
!RegEx Regex does NOT match "/ERROR/"
#comparison Extract + compare two values See below

#comparison: Extract values with regex, compare numerically:

{
  "rule": "/Available: (\\d+)/",
  "ruleB": "/Total: (\\d+)/",
  "eval": "#comparison",
  "evaluator": ">=",
  "severity": "error"
}

Evaluators: =, !=, <, >, <=, >=, % (percentage)

Flags: case: true = case-INSENSITIVE (confusing name), global: true, multiline: true (RegEx only)

Run a Command Template

Standalone:

POST /mop/RunCommandTemplate
{
  "template": "Port_Turn_Up_Pre_Check",
  "variables": {"interface": "GigabitEthernet0/1", "vlan_id": "100"},
  "devices": ["IOS-CAT8KV-1"]
}

In a workflow (MOP.RunCommandTemplate task):

{
  "incoming": {
    "template": "$var.job.templateName",
    "variables": "$var.job.templateVariables",
    "devices": "$var.job.devices"
  },
  "outgoing": {
    "mop_template_results": null
  }
}

Response Shape

{
  "all_pass_flag": true,
  "result": true,
  "name": "Port_Turn_Up_Pre_Check",
  "commands_results": [
    {
      "raw": "show interface <!interface!>",
      "evaluated": "show interface GigabitEthernet0/1",
      "all_pass_flag": true,
      "device": "IOS-CAT8KV-1",
      "response": "...command output...",
      "result": true,
      "rules": [{"rule": "line protocol is", "eval": "contains", "result": true}]
    }
  ]
}

Update a Command Template

POST /mop/updateTemplate/{mopID}

mopID is the template name (URL-encoded). Body is {"mop": {...}} — full replacement, include ALL fields.

Analytic Templates (Pre/Post Comparison)

POST /mop/createAnalyticTemplate
{
  "name": "Interface_Change_Validation",
  "os": "cisco-ios",
  "passRule": true,
  "prepostCommands": [
    {
      "preRawCommand": "show interface GigabitEthernet0/1",
      "postRawCommand": "show interface GigabitEthernet0/1",
      "passRule": true,
      "rules": [
        {
          "type": "matches",
          "preRegex": "/line protocol is (\\w+)/",
          "postRegex": "/line protocol is (\\w+)/",
          "evaluator": "="
        }
      ]
    }
  ]
}

In a workflow (MOP.runAnalyticsTemplate task):

{
  "incoming": {
    "pre": "$var.preCheckTaskId.mop_template_results",
    "post": "$var.postCheckTaskId.mop_template_results",
    "analytic_template_name": "Interface_Change_Validation",
    "variables": {}
  },
  "outgoing": {"analytic_result": null}
}

Testing & Debugging

Start a Job

POST /operations-manager/jobs/start
{
  "workflow": "My Workflow Name",
  "options": {
    "description": "Test run",
    "type": "automation",
    "variables": {"deviceName": "IOS-CAT8KV-1"}
  }
}

Response: {"message": "...", "data": {"_id": "jobId", "status": "running"}}

Check Job Status

GET /operations-manager/jobs/{jobId}

Response wrapped in {message, data, metadata}:

  • data.status — "running", "complete", "error", "canceled"
  • data.variables — all job variables including outputs
  • data.error — array of error objects on failure

Debug Failed Jobs

  1. GET /operations-manager/jobs/{jobId} — check data.status
  2. If "error", read data.error[] — each has task (ID) and message.IAPerror.displayString
  3. Identify the failing task ID, check its metrics.finish_state

Common failures:

Symptom Cause Fix
“Method not found” validation error Task name doesn’t exist Search tasks.json
“No available transitions” Missing error transition Add "state": "error" transition
$var resolves to literal string Non-hex task ID or nested object Check task IDs, use merge
“Cannot find workflow” childJob ref broken after project move Update workflow field with @projectId: prefix
Schema validation error Wrong/missing fields Check task-schemas.json
Adapter error Wrong app name or adapter down Check apps.json and GET /health/adapters
“No config found for Adapter: X” app field uses adapter instance name instead of type name app/locationType must be the type from apps.json (e.g., EmailOpensource), not instance name (e.g., email). Instance name goes in adapter_id.
Silent data mismatch Field type doesn’t match schema (string vs array) Check task-schemas.json — pass arrays for array fields, numbers for number fields
Same type-mismatch warning recurs after a “fix” that changed task type but not data shape $var Resolution Rule violated (object/array passed as static value) Don’t swap task types — build the value through merge/makeData per the $var Resolution Rules section, then re-validate
Workflow built via restCall against Itential’s own internal API (e.g., /gateway_manager/v1/...) Skipped searching tasks.json for a native task before reaching for restCall Search tasks.json filtered by app (e.g., GatewayManager, InventoryManager) — a native task almost always exists
“Schema validation failed on must have required property ‘X’” Missing field in adapter body Add the field to the merge task building it
“Referenced job variable: undefined” merge’s data_to_merge used "value" instead of "variable" Change to "variable" — see Variable Syntax Reference
Job stuck in "running" indefinitely No error transition on the failing task Add a "state": "error" transition

Fix locally, PUT to update, re-run — don’t recreate; updating preserves the ID.

Standalone Test Endpoints

See Task Endpoint Patterns (Standalone Testing) above (in Guides) for the full list of tasks with standalone REST endpoints and the dummy-job_id trick.

Updating Assets (Edit Locally, PUT to Update)

Asset Create Update Delete
Workflow POST /automation-studio/automations PUT /automation-studio/automations/{id} with {"update": {...}} DELETE /workflow_builder/workflows/delete/{URL-encoded-name} (by name, not ID)
Template POST /automation-studio/templates PUT /automation-studio/templates/{id} with {"update": {...}} DELETE /automation-studio/templates/{id}
Command Template POST /mop/createTemplate POST /mop/updateTemplate/{name} with {"mop": {...}} (full replacement) —

Pre-flight validate before every create or update — use the deep validator (6.5.2+):

POST /workflow_engine/workflows/validate
{"asset": {...workflow document...}}
→ {"isValid": true|false, "errors": [], "warnings": []}

isValid is exactly errors.length === 0 — warnings never affect it. This is the default pre-flight gate; it is stateless and safe to call as often as needed. Prefer it over POST /automation-studio/workflows/validate ({"workflow": {...}} → {errors, warnings}, no isValid), which misses errors this one catches.

The endpoint can return HTTP 500 instead of a normal body. A manual-view task (e.g. ViewData) with type set to anything other than "manual" returns {"message":"An unknown error occurred...", "data":{"error":"Cannot destructure property 'input' of 'undefined'..."}} at HTTP 500. Always check for this shape before parsing isValid/errors/warnings. Fix the task’s type first if you hit it.

What this endpoint catches, by category:

  • Errors (block isValid): non-hex/reserved task IDs, workflow-level required fields, per-task required fields (including description, required on every task), adapter app checked against the platform’s registered adapter types ("No config found for Adapter: X"), adapter adapter_id checked against live registered adapter instances ("X is not a configured adapter within the platform" — exempts $var/task-ref values), adapter/application method incoming and outgoing field-name mismatches, transition type/state enum values, dangling/missing transition entries, cycles in standard transitions, created_by/last_updated_by object-vs-string shape on import documents.
  • Warnings only (isValid stays true — inspect warnings[] separately): nested $var inside an object/array incoming value (see the allow-list exceptions in the $var Resolution Rules section below), static-value type mismatches ("should be of type X but is of type Y"), enum-typed static values with a non-enumerated value, dangling/out-of-order job-variable and task-to-task $var references.
  • Not checked at all — keep doing these manually: missing error/failure transitions (only a missing success path is flagged), evaluation.operator invalid values like "regex" (the closed enum is only wired to the standalone evaluation-test endpoints, not to a workflow document’s embedded evaluation task), merge/childJob "value" vs "variable" key-naming mistakes, wrong canvasName, incomingRefs cache staleness after PUT, adapter response-shape assumptions, projects/import semantics (different service).

A non-empty warnings array is not “good enough to ship.” errors blocks a job from starting; warnings (e.g., "input should be of type string but is of type object", "outputType is of type enum but got non-enumerated value...") means a task’s static data won’t behave the way you intended even though the workflow will technically save and run. Treat any non-empty warnings array as a build defect to fix before marking the component delivered — don’t report a workflow as built/delivered just because the save call returned a 200. If the same warning reappears after a “fix,” you’ve patched the symptom, not the cause — re-check the $var Resolution Rules section below before trying another variation.

Validate every asset with its own validator before creating, updating or importing it, and read both errors and warnings. Each asset type has its own endpoint and request shape; sending a document to the wrong one, or wrapping a template or transformation in asset, returns misleading errors:

Asset Call Request body
Workflow POST /workflow_engine/workflows/validate {"asset": <workflow document>}
JSON form POST /json-forms/validate {"asset": <form document>}
Template POST /automation-studio/templates/validate the template document itself — no asset wrapper
Transformation POST /transformations/validate the transformation document itself — no asset wrapper

Each returns {"isValid": true|false, "errors": [], "warnings": []}. For any other asset type (e.g., MOP command templates), look up its validator in openapi.json; don’t guess.

Validator quirks:

  • An exported template stores createdBy and lastUpdatedBy as user objects; the template validator rejects them (“must be a MongoDB ObjectId”). Replace each with the user’s _id before validating.
  • The template validator rejects an empty or whitespace-only command or description, though import accepts them. Give both real values.
  • The workflow validator can flag query decorators as errors when they aren’t defects — see “Validator vs runtime” above. Don’t remove the decorators to silence it.
  • The form validator catches a missing top-level nodeId but not one on a field nested in an array field. Check nested fields yourself (see the itential-json-forms skill).

Never author a large workflow JSON payload as an inline shell string. Write it to a file with a proper file-editing tool, or if constrained to bash, generate it via python3 -c "..." with json.dump() — never an interactive heredoc with embedded Jinja/markdown/emoji. If a payload fails to parse, regenerate it cleanly rather than patching it byte-by-byte with sed.

Workflow rename:

POST /workflow_builder/workflows/rename
{"workflow": {...full doc...}, "newName": "New Workflow Name"}

Renames in-place without recreating. Use instead of appending [Fixed] suffixes.

Fetch → modify → PUT → re-verify, in one contiguous chain — never PUT from a locally-cached copy. If you fetched a workflow/project/template earlier in the session and it may have changed since (you or something else edited it), a PUT built from that stale local file will silently revert any change made in between — the PUT succeeds, but it overwrites the newer state with the old one. Always: fetch fresh immediately before modifying, PUT, then GET again immediately after to confirm the change actually landed. This is easy to trigger across a multi-step debugging session where the same asset is fetched and modified several times.


Workflow Patterns

Error Handling: Try-Catch

In child workflows: catch errors with newVariable to set a status flag:

task --success--> newVariable("taskStatus" = "success") -> workflow_end
task --error--> newVariable("taskStatus" = "error") -> workflow_end

In parent workflows: after childJob, extract and check:

childJob -> query (extract taskStatus from job_details) -> evaluation (== "success"?)
  |-- success -> continue
  |-- failure -> handle error

Manual Tasks (Human-in-the-Loop)

Pick the task by what the person has to do:

Use case Task Outgoing
Approve/reject one thing, or acknowledge WorkFlowEngine ViewData none — the button sets the finish state (success/failure) for the next transition
A custom form: choose options, tick items, type values WorkCenter InteractiveHTML export: each form input’s value, keyed by its name
Act on each row of a list WorkCenter QuickForm rows: the rows, each with the person’s input under actionColumnKey

ViewHTML only displays HTML and returns nothing; use InteractiveHTML when you need values back. InteractiveHTML and QuickForm are type: "manual", app: "WorkCenter", with view /work-center/task/InteractiveHTML and /work-center/task/QuickForm. Take their other fields from GET /workflow_builder/tasks/list (or multipleTaskDetails) rather than copying the ViewData shape below, and validate before saving. Their inputs and examples are under “Human approval” in /flowagent.

ViewHTML — renders an HTML string in a modal for operator review. Requires specific fields or it becomes a draft workflow:

{
  "name": "ViewHTML", "canvasName": "ViewHTML",
  "location": "Application", "locationType": null, "app": "WorkFlowEngine",
  "type": "manual", "displayName": "Tools",
  "view": "/workflow_engine/task/ViewHTML",
  "taskVersion": 2, "hostApp": "@itential/app-operations_manager",
  "variables": {
    "incoming": {
      "header": "Report Title",
      "body": "$var.job.html_output",
      "variables": "",
      "btn_success": "Acknowledge",
      "btn_failure": ""
    },
    "outgoing": {}
  },
  "actor": "Pronghorn"
}

view, taskVersion: 2, and hostApp are all required — omitting any one causes “Manual Tasks require ‘view’ key” draft error.

Read a live ViewData example from the Cisco IOS upgrade workflow — it shows makeData → ViewData → success/failure branches in production:

jq '[.components[].document.tasks // {} | to_entries[] | select(.value.name == "ViewData")] | first | .value' \
  assets/helpers/assets/vendor-cisco-ios.json

Three rules that cause draft validation errors if missed:

  1. view is a top-level field (sibling of name, type, app) — NOT inside variables. Missing it → "Manual Tasks require 'view' key with path to task view".
  2. incoming.variables MUST be present (value can be {} if unused). Missing it → "Input: 'variables' is not defined in task model".
  3. displayName must be "Tools" and actor must be null (no actor field) on manual tasks.

Note: production assets include "error": "" and "decorators": [] in the variables block on ViewData tasks — these are added by Studio on export and are harmless. You do not need to add or remove them.

{
  "name": "ViewData",
  "canvasName": "ViewData",
  "location": "Application",
  "app": "WorkFlowEngine",
  "displayName": "Tools",
  "type": "manual",
  "view": "/workflow_engine/task/ViewData",
  "variables": {
    "incoming": {
      "header": "Approval Required",
      "message": "Review and approve.",
      "body": "$var.job.dataToReview",
      "variables": "$var.job.dataToReview",
      "btn_success": "Approve",
      "btn_failure": "Reject"
    },
    "outgoing": {}
  },
  "groups": []
}

ViewHTML (Manual Task — HTML Display)

Use ViewHTML when you need to display formatted HTML to an operator during a workflow — for reports, tables, or styled summaries. Same manual task rules as ViewData apply.

Read a live ViewHTML example:

jq '[.components[].document.tasks // {} | to_entries[] | select(.value.name == "ViewHTML")] | first | .value' \
  assets/helpers/assets/vendor-cisco-ios.json

Key differences from ViewData:

  1. view is /workflow_engine/task/ViewHTML
  2. body is a raw HTML string — use inline CSS (no <style> blocks), <!var!> syntax for variable substitution
  3. incoming.variables supplies the <!var!> placeholders in the HTML: a static object {"varName": "value"}, or a whole-field reference to an object built upstream ("$var.job.x", or with Enable Query "$var.job.x#/body" plus its decorator). A $var placed inside the static object is not resolved
  4. displayName: "Tools", no actor field (same as ViewData)
{
  "name": "ViewHTML",
  "canvasName": "ViewHTML",
  "location": "Application",
  "app": "WorkFlowEngine",
  "displayName": "Tools",
  "type": "manual",
  "view": "/workflow_engine/task/ViewHTML",
  "variables": {
    "incoming": {
      "header": "Report",
      "body": "<h2>Status: <!status!></h2><p>Device: <!device!></p>",
      "variables": {
        "status": "$var.job.deviceStatus",
        "device": "$var.job.deviceName"
      },
      "btn_success": "Continue",
      "btn_failure": "End"
    },
    "outgoing": {}
  },
  "groups": []
}

autoApprove Pattern

Use an evaluation task to conditionally skip manual approval:

evaluation (autoApprove == true?)
  |-- success -> skip to next task (auto-approved)
  |-- failure -> ViewData (human reviews and approves/rejects)

The workflow accepts an autoApprove boolean input. When true, skips the manual step. Useful for CI/CD pipelines that run unattended vs interactive operator sessions.

Revert Transitions (Retry Loops)

Use "type": "revert" transitions to go backward for retry scenarios:

renderTemplate -> viewConfig (approve/reject)
  |-- success -> pushConfig -> evalSuccess
  |                             |-- success -> end
  |                             |-- failure -> viewError (retry/abort)
  |                                             |-- success (retry) --revert--> renderTemplate
  |                                             |-- failure (abort) -> end
  |-- failure (reject) --revert--> renderTemplate

The revert transition moves execution back to a previous task, allowing the user to fix inputs and retry.

Modular Workflow Design

  • Build each child workflow independently testable via jobs/start
  • Use childJob with data_array + loopType: "parallel" to fan out
  • Check for existing workflows before building new ones
  • Keep all asset JSON locally — edit locally, PUT to update

Network Device Config Pattern

  1. MOP command templates for validation checks only (show commands + rules)
  2. Jinja2 templates to generate configuration
  3. Push config via existing workflow or adapter task — ask the engineer
  4. Test CLI commands on the actual device BEFORE building workflows

Variable Syntax Reference

Context Syntax Example
Jinja2 templates {{ var }} interface Vlan{{ vlan_id }}
Command templates (MOP) <!var!> show interface <!interface!>
makeData input <!var!> {"name": "<!name!>"}
Workflow variable refs $var.job.x or $var.taskId.x $var.job.deviceName
childJob variable refs {"task":"job","value":"varName"} {"task":"static","value":["a"]}
merge/evaluation refs {"task":"job","variable":"varName"} {"task":"static","variable":"success"}

childJob uses "value". merge/evaluation use "variable". Do NOT mix them.


API Response Shapes

Endpoint Shape
POST /operations-manager/jobs/start {message, data: {_id, status}}
GET /operations-manager/jobs/{id} {message, data: {status, variables, error}}
POST /automation-studio/projects {message, data: {_id, name}}
POST /automation-studio/automations {created: {_id, name}, edit: "..."}
POST /automation-studio/templates {created: {_id, name}, edit: "..."}
GET /automation-studio/workflows {items: [...], skip, limit, total}
GET /automation-studio/templates {items: [...], skip, limit, total}

Adapter Response Shapes

Adapters transform upstream API responses. Don’t assume the native API’s response structure. For example, ServiceNow’s Table API returns result.sys_id, but the Itential adapter flattens it to response.id. Always verify by calling the adapter directly or checking openapi.json.

Adapter URI Prefix

genericAdapterRequest auto-prepends the adapter’s base_path to uriPath. Don’t include /api/v1 in uriPath. Use genericAdapterRequestNoBasePath to bypass.


Gotchas

Pre-flight scan list — read this before every project import and job start. This list is intentionally redundant with the body sections. The repetition is deliberate: scanning a flat list before submitting catches mistakes that are easy to miss when building task by task.

To add a new finding: put the detail in the relevant body section first, then add a one-liner here pointing to it.

Projects

  1. Use POST /projects/import to create projects with all assets atomically — avoids broken childJob refs, project-locking issues, and intermediate state. Pre-compute _id so childJob @projectId: refs can be wired before push.
  2. Avoid create + move pattern — moving assets renames them with @projectId: prefix but does NOT update internal references (childJob workflow fields, template names).
  3. Import format differs from create — OMIT encodingVersion from workflow documents (causes silent component failure). Workflow created_by has NO _id but has firstname, inactive, sso. Project createdBy HAS _id.
  4. Component type is mopCommandTemplate — not mop.
  5. Members PATCH is full replacement — include ALL members or omitted ones are removed.
  6. Import sets the OAuth service account as project owner — not the UI user. PATCH membership immediately after import (Phase 3, not Phase 6) or the spec engineer is locked out.
  7. accessControl in PATCH body is silently ignored — API returns 200 but the field is a no-op. Always use the members array format ([{type, reference, role}]). 7a. Component reference formats differ — a workflow’s is a uuid; a template, transformation or JSON form’s is a 24-character hex id. A uuid on a template fails the import with Argument passed in must be a string of 12 bytes or a string of 24 hex characters, which names nothing. Check metadata.failedComponents after every import.

Workflows

  1. canvasName must come from tasks.json — some differ from method name: arrayPush→push, stringConcat→concat. Wrong canvasName causes silent $var failures.
  2. Task IDs must be hex [0-9a-f]{1,4} — non-hex causes silent $var failure.
  3. Validation errors = draft workflow that cannot be started. Run POST /workflow_engine/workflows/validate (deep validation, prefer over /automation-studio/workflows/validate) before every create or update, and check warnings[] too — isValid ignores them. 10a. Each asset type has its own validator and request shape — templates and transformations take the document itself, with no asset wrapper. See “Validate every asset with its own validator”.
  4. $var inside nested objects doesn’t resolve — use merge/makeData/query to build the object first. Exception: GatewayManager.runService/runServiceStatic (params), GatewayManager.runCode (data), AgentSessionManager.runAgent (inputs) resolve one level deep into that one specific key. 11a. Enable Query needs a suffix AND a decorator — $var.x#/path alone runs fine but Studio shows no query, and a Studio save then strips the #/path. The pointer depends on the field shape (merge and childJob differ); on evaluation put query/rightQuery on the evaluation object, never inside the operand. Verify with assets/helpers/enable_query.py check — see Enable Query.
  5. stringConcat does not resolve $var inside stringN arrays — values stored as literal strings. Use merge → makeData with <!var!> placeholders instead.
  6. Every adapter/external task needs an error transition — without one, errors cause “Job has no available transitions” and the job gets stuck forever. POST /workflow_engine/workflows/validate does not check for this — it only flags a missing success path. 13a. Scenario input must be an object ({} when empty), and tr_id must be an id the platform returned — create transformations before the workflows that call them. See Workflow Structure and ### transformation.
  7. JSON can’t have duplicate keys — if success and error both go to workflow_end, route error to an intermediate newVariable task first. 14a. Every workflow document needs "canvasVersion": 3 — see the Step 9 checklist. Without it, validation fails and import stores the workflow as canvas version 1.

Utility Tasks

  1. merge uses "variable", childJob uses "value" — don’t mix them. Using "variable" in childJob causes undefined.indexOf() at job start (P6.4.0+).
  2. merge requires at least 2 items — 1 item silently returns null.
  3. childJob actor MUST be "job", task MUST be "", job_details outgoing MUST be null.
  4. childJob variables use {"task","value"} NOT $var — $var strings inside variables cause an indefinite hang at runtime.
  5. evaluation MUST have both success AND failure transitions — missing one silently drops the job.
  6. forEach last body task transition must be empty {} — do NOT connect it back to forEach.
  7. push/pop/shift take variable NAME as a plain string — "myArray" not "$var.job.myArray".
  8. newVariable value with $var stores the literal string — use merge + query to build dynamic values.
  9. makeData variables must be a resolved object — use merge first, then pass $var.taskId.merged_object.
  10. Adapter task result is always an object — never a primitive. When the upstream API returns a simple string (e.g., Infoblox _ref), it’s at result.response. Passing raw result in a string context produces [object Object].
  11. childJob cannot resolve project-scoped workflow names — "@<projectId>: <name>" and the bare name both fail with "Cannot find workflow ...", even from a calling workflow in the same project. Inline the target task(s) instead; there’s no known childJob workaround. See ### childJob for how this was confirmed.
  12. A merge reading a childJob’s output uses "variable", like every other merge item — {"task": "<childJobId>", "variable": "job_details"}. With "value" the workflow creates and runs but the item resolves to null. See ### childJob.
  13. push’s incoming must omit job_id (even though tasks.json lists it) and its outgoing must declare job_variable_value (not an empty {}) — both cause the same generic create/update failure as #26 if wrong.
  14. parse’s real fields are text/textObject, not stringToParse/result.
  15. objectToString’s replacer is a property whitelist — replacer: [] silently produces {} for every input, no error, no warning. replacer/space are both optional (required: false); if you don’t need them, OMIT the keys entirely rather than passing null (which trips a validate-time type warning) or []/0 (which is schema-valid but semantically means “include zero properties” — the empty-whitelist regression). This is a genuine silent-data-loss bug, not cosmetic: the task’s own stringified output reads as "{}" with no error while $var.job.<sourceVar> is fully populated the whole time. If you see objectToString “losing” data that a job’s final variable dump shows was correct, check replacer first before suspecting a timing/race issue in whatever produced the source object.
  16. $var does not resolve when written as an element inside an inline array literal — "groups": ["$var.job.myGroup"] sends the literal unresolved string, not the resolved value (confirmed via A/B test: ["academy"] worked, ["$var.job.groupVar"] sent the literal 21-character string and the API rejected it as an unknown group name). Same family as “no $var inside nested object values,” but applies to plain arrays of primitives too. Fix: build the array with newVariable (value: []) + push (one item at a time), then reference the whole array as a single top-level $var.job.<name> — a bare job-variable reference resolves fine.
  17. InventoryManager.getNodesByInventory’s params field is required at runtime even though its live schema (multipleTaskDetails?dereferenceSchemas=true) doesn’t mark it required: true. Omitting it fails job start with "Cannot find match for input: \"params\" from model". Always pass "params": {} explicitly for this task, regardless of what the schema implies.
  18. InventoryManager.getInventoryByIdentifier surfaces “not found” as a task-level error transition, not a success with an error-shaped body (confirmed with an isolated probe: HTTP 404 → job status: error, task’s outgoing.response stays unset). Branch on the task’s own success/error transitions for existence checks — no evaluation task needed to inspect a response field.
  19. Put an evaluation query on the evaluation item (query / rightQuery), never inside the operand — {"task":"job","variable":"src","query":"body.name"} is ignored at runtime for every path form (operand resolves to the whole object, the comparison silently fails) and Studio shows nothing for it. Use {"query":".body.name","operand_1":{"task":"job","variable":"src"},"operator":"==","operand_2":{...},"rightQuery":""}. If a query still has no effect, fall back to a standalone query task + {"task":"<queryTaskId>","variable":"return_data"} operand ref. See the ### evaluation section.
  20. query’s pass_on_null: false does not reliably trigger failure for an out-of-bounds array index (e.g. results[0] on an empty []) — observed return_data: null but success transition still taken. Don’t rely on the failure transition to catch this case; check the extracted value explicitly downstream. See ### query.

Templates

  1. Template group cannot be empty or whitespace-only — causes a silent rejection.
  2. TextFSM templates may contain control characters that break jq — use Python with a control-char strip when parsing them.
  3. renderJinja2 inline template with \n breaks parse — static values store literal \n characters, causing parse to fail with “Expected property name or ‘}’ in JSON at position 1”. Fix: write single-line templates. 37a. Template data must be a non-blank string, never {} — missing or {} fails that component on import while the project still reports success; check metadata.failedComponents. See ## Templates.

MOP

  1. Missing variable = skip = PASS (not fail) — if a variable isn’t passed, the rule is skipped and the command auto-passes. Always verify variables are passed correctly.
  2. case: true = case-INsensitive — the name is backwards. Easy to wire the wrong behavior.
  3. Eval types are case-sensitive — "RegEx" not "regex". Wrong casing silently fails.
  4. Empty rules = auto-pass — a command with no rules always passes. Add at least one rule to validate output.
  5. MOP update is full replacement — include ALL fields or omitted ones are lost.
  6. MOP is read-only — never use it to push config. Use GatewayManager.sendConfig for config push; AGManager.itential_cli (Gateway4) only inside an existing Gateway4 workflow.

API / Endpoints

  1. Adapter app must come from apps.json — NOT tasks.json. Names can differ completely (e.g., ServiceNow vs Servicenow). Wrong app causes “No config found for Adapter” at runtime.
  2. legacyWrapper: false on Operations Manager manual triggers — default true wraps all form values under formData, breaking variable mapping to workflow inputs.
  3. status: complete doesn’t mean CLI commands succeeded — always check stdout for actual command output and errors.
  4. Endpoint base paths differ — task catalog at /workflow_builder/tasks/list, schemas at /automation-studio/multipleTaskDetails (NOT /workflow_builder/multipleTaskDetails).
  5. evaluation operator is a closed enum — only contains, !contains, <, <=, >, >=, ==, != exist. Any other string silently returns false with empty outgoing and no error message. Validate against this list before wiring.
  6. contains uses regex, not substring matching — operand_2 is interpreted as a regex pattern. Escape metacharacters ((, ), ., [, ], ?, +, *, |) in literal values: 9\.2\(4\) not 9.2(4). Test with POST /workflow_engine/runEvaluationGroups before wiring.
  7. API PUT does not regenerate incomingRefs for existing task changes — evaluation operand literals resolve to null after PUT. Broader symptom: entire workflow hangs after workflow_start (status: running forever). Fix: open in Studio and save. If still failing, recreate via fresh POST — more PUTs won’t fix it.
  8. $var.<taskId>.<out> does not resolve inside nested forEach bodies — use $var.job.<varName> for any variable referenced inside a nested loop body.
  9. Workflow delete endpoint — DELETE /workflow_builder/workflows/delete/{URL-encoded-name} deletes by name, returns 200 with deleted doc. DELETE /automation-studio/automations/{id} does NOT exist (404). Always export the project before deleting anything.
  10. Task outgoing writes directly to job var — "outgoing": {"result": "$var.job.myVar"} works on any task and is more reliable than a separate newVariable copy step (written at execution time, bypassing incomingRefs cache).
  11. GatewayManager "failed to parse start_time" = device unreachable — this IAG error ("failed to parse start_time for command 0: failed to parse timestamp string ''") means the device is offline, unreachable, or auth failed. The timestamp complaint is misleading — the session never opened. It is NOT a workflow bug. Guard with an evaluation checking whether the response contains a result key; if not, route to a skip handler and continue.
  12. Project component refresh — mode: "copy" creates a new project-scoped UUID that immediately diverges from the standalone. To refresh: DELETE each old component, POST fresh, then update any Operations Manager automation componentId via PATCH /operations-manager/automations/{id}.

Workflow Design Heuristics

  1. Always use a local venv for Python — python3 -m venv .venv && source .venv/bin/activate before any Python scripts during the build.
  2. Search tasks.json before designing any sub-workflow — a purpose-built platform task may already exist for the intent (filter, inventory, tag, etc.). Server-side is always better than a forEach + evaluation chain.
  3. Prefer server-side filtering over client-side when available — fetching the full collection and filtering in a forEach adds unnecessary iterations. Check for a filtered-fetch task first.
  4. Propose decomposition when a workflow exceeds ~20 tasks — extract inner iteration bodies into reusable child workflows.
  5. DRY check on sibling workflows — if building multiple similarly-named workflows, compare task graphs. Identical graphs → propose one generic workflow, not N clones.
  6. NEVER wire a Configuration Manager remediation task — see assets/AGENTS.md Rule 25 for the full prohibited-task list and the config-push alternative.

API / Endpoints (continued)

  1. POST /workflow_builder/workflows/save: set scenarios: [], errors: [], warnings: [] and canvasVersion: 3 explicitly on a from-scratch document — without canvasVersion, validation returns isValid: false and import stores the workflow as canvas version 1; jobs still start. A jobs/start crash with "Cannot convert undefined or null to object" has another cause (see the adapter list/search tasks note). Prefer POST /automation-studio/automations for creation. See ### Workflow Structure.
  2. Admin-bypass project PATCH (/automation-studio/admin/projects/{id}) silently drops components not listed in the same request — even a members-only PATCH can delete an unrelated asset outright. Always diff components after using this endpoint. See ### Update membership.
  3. Adapter list/search tasks need every declared query parameter present (even as "") — partial incoming validates but crashes jobs/start the same generic way. Copy the full block from a wired example. See ### Get Full Task Schemas.
  4. Check each query parameter’s declared type before assigning a plain string — some are array-typed even when they read like single-value filters (e.g. NetBox ipam_ip_addresses_list’s parent). See ### Get Full Task Schemas.
  5. DELETE-with-body adapter tasks (e.g. *_bulk_destroy) fail at the gateway/proxy layer, not the adapter — loop single-item _destroy calls instead. See ### Get Full Task Schemas.

Helper Templates

Two separate concerns — don’t mix them:

  • API wrappers (project, workflow, template, form creation) → use assets/helpers/create/ scaffolds below. These are POST body wrappers — correct structure, required fields.
  • Task JSON inside a workflow → extract from assets/helpers/assets/ using jq (see Guide 1 STOP block). Do NOT use assets/helpers/create/ files for task bodies — they are scaffold stubs, not task examples.

Scaffolds — start from these

Read these first. They have the correct wrapper, required fields, and structure.

When you need to… Read this helper Then POST to
Create a project assets/helpers/create/create-project.json POST /automation-studio/projects
Create a workflow assets/helpers/create/create-workflow.json POST /automation-studio/automations
Create a Jinja2 template assets/helpers/create/create-template-jinja2.json POST /automation-studio/templates
Create a TextFSM template assets/helpers/create/create-template-textfsm.json POST /automation-studio/templates
Create a MOP command template assets/helpers/create/create-command-template.json POST /mop/createTemplate
Update a MOP template assets/helpers/update/update-command-template.json POST /mop/updateTemplate/{name}
Create a JSON form (static dropdowns) assets/helpers/create/create-json-form.json POST /json-forms/forms — see itential-json-forms skill
Create a JSON form (REST-bound or cascading dropdowns) assets/helpers/create/create-json-form-rest-bound.json POST /json-forms/forms — see itential-json-forms skill
Create a JSON form (array-of-objects field) assets/helpers/create/create-json-form-array-field.json POST /json-forms/forms — see itential-json-forms skill
Create an Ops Manager automation assets/helpers/create/create-ops-manager-automation.json POST /operations-manager/automations
Create a manual trigger (with form) assets/helpers/create/create-ops-manager-trigger-manual.json POST /operations-manager/triggers — legacyWrapper MUST be false
Create a scheduled trigger assets/helpers/create/create-ops-manager-trigger-schedule.json POST /operations-manager/triggers
Import a project (atomic) assets/helpers/create/import-project.json POST /automation-studio/projects/import
Add assets to a project assets/helpers/operations/add-components-to-project.json POST /projects/{id}/components/add
Update project membership assets/helpers/update/update-project-members.json PATCH /projects/{id}

Task templates — extract from asset projects

Do not write task JSON from scratch. For every task type, extract a real example from an asset project and adapt it. Use the jq commands below — they work against the project files in assets/helpers/assets/.

# Adapter task (ServiceNow, Infoblox, etc.)
jq '[.components[].document.tasks // {} | to_entries[] | select(.value.location == "Adapter")] | first | .value' \
  assets/helpers/assets/vendor-servicenow.json

# Application task (WorkFlowEngine — getTime, newVariable, query, evaluation, transformation, merge, makeData)
jq '[.components[].document.tasks // {} | to_entries[] | select(.value.app == "WorkFlowEngine" and .value.name == "TASK_NAME")] | first | .value' \
  assets/helpers/assets/itential-platform-configuration-management.json

# childJob
jq '[.components[].document.tasks // {} | to_entries[] | select(.value.name == "childJob")] | first | .value' \
  assets/helpers/assets/vendor-cisco-ios.json

# evaluation / branching
jq '[.components[].document.tasks // {} | to_entries[] | select(.value.name == "evaluation")] | first | .value' \
  assets/helpers/assets/vendor-cisco-ios.json

# transformation (JST)
jq '[.components[].document.tasks // {} | to_entries[] | select(.value.name == "transformation")] | first | .value' \
  assets/helpers/assets/vendor-netbox.json

# RunCommandTemplate / viewTemplateResults / reattempt / runTemplatesDiff (MOP tasks)
jq '[.components[].document.tasks // {} | to_entries[] | select(.value.name == "TASK_NAME")] | first | .value' \
  assets/helpers/assets/itential-platform-configuration-management.json

# sendConfig (config push via Gateway5) and the runCode task that builds its config and inventory
jq '[.components[].document.tasks // {} | to_entries[] | select(.value.name | test("^(sendConfig|runCode)$"))] | map(.value)' \
  assets/helpers/assets/vendor-arista-eos.json

# ViewData / ViewHTML (manual tasks)
jq '[.components[].document.tasks // {} | to_entries[] | select(.value.name == "ViewData")] | first | .value' \
  assets/helpers/assets/vendor-cisco-ios.json

Key fields to verify after extracting (see body sections above for full rules per task type):

Task type Must-check fields
Adapter app/locationType from apps.json (not tasks.json), adapter_id, error transition
childJob actor: "job", task: "", variables use {"task","value"} not $var
evaluation evaluation_groups[], all_true_flag, both success AND failure transitions
transformation tr_id, variableMap keys match transformation’s incoming schema
ViewData / ViewHTML view is top-level (not inside variables), displayName: "Tools", no actor, no error/decorators
Manual tasks (any) type: "manual", taskVersion: 2, hostApp required

Reference workflows — read from asset projects

Real, production-tested workflows. Use the jq commands to extract and study them before building.

Pattern Asset file jq filter
Adapter workflow: merge → create → query → update + error handling assets/helpers/assets/vendor-servicenow.json select(.document.name | test("Create Change"))
childJob orchestrator + evaluation branching assets/helpers/assets/vendor-cisco-ios.json select(.document.name | test("IOS Upgrade"))
childJob loop with data_array (parallel/sequential) assets/helpers/assets/vendor-cisco-ios.json select(.document.name | test("Upgrade|Runner"))
Config push: renderJinja → approval ViewData → sendConfig → per-node success check assets/helpers/assets/vendor-arista-eos.json select(.document.name | test("Push Config"))
Pre/post MOP check: RunCommandTemplate → viewTemplateResults → evaluation → reattempt assets/helpers/assets/itential-platform-configuration-management.json select(.document.name | test("Command Template Runner"))
IPAM CRUD (adapter + transformation + error) assets/helpers/assets/vendor-infoblox-nios-ddi.json select(.document.name | test("Create Network|Assign Next"))
ITSM ticket + update (ServiceNow) assets/helpers/assets/vendor-servicenow.json select(.document.name | test("Create Incident"))
LCM action workflow (must output instance) lcm/lcm-vxlan-fabric-services-project.json select(.document.name | test("Create|Delete")) — note: this file uses .data.project.components[]
Email/notification assets/helpers/assets/itential-platform-email.json select(.document.name | test("Email"))
Per-item transform loop replaced by runCode + Enable Query (zero/near-zero WorkFlowEngine utility tasks) assets/helpers/assets/netbox-inventory-sync-runcode-enablequery.json select(.document.name | test("runCode")) — 2 real workflows: one calling an external script, one using only native app tasks, both with the same runCode device-mapping pattern
# General pattern to read any workflow by name from an asset project:
jq '[.components[] | select(.type=="workflow") | select(.document.name | test("PATTERN"; "i"))] | first | .document | {name:.name, tasks:.tasks, transitions:.transitions}' \
  assets/helpers/assets/ASSET_FILE.json

# For the LCM project (different wrapper):
jq '[.data.project.components[] | select(.type=="workflow") | select(.document.name | test("PATTERN"; "i"))] | first | .document | {name:.name, tasks:.tasks, transitions:.transitions}' \
  assets/helpers/assets/lcm/lcm-vxlan-fabric-services-project.json