Workflow Automation

The host platform’s model is rigid: every Apify capability must be expressed as one of trigger, action (create), or search (read). Decide the mapping before writing any code.

1. Map Apify to the host’s trigger / action / search model

Reads of the last run or stored data are searches, because that is the host’s mechanism for “find an existing record.” Run-finished is a webhook trigger. Actor/Task selection is a hidden trigger that feeds dropdowns - reusing one hidden trigger across multiple actions keeps the surface DRY.

2. Canonical capability matrix

The complete set of operations a workflow integration should surface. Treat this as the floor, not a menu to trim without reason.

Triggers

• Actor run finished (webhook; user picks terminal statuses: SUCCEEDED / FAILED / TIMED-OUT / ABORTED)
• Task run finished (same shape, scoped to a Task)

Actions (creates)

• Run Actor - fire an Actor run
• Run Actor and get dataset items - fire and return results inline
• Run Task - fire a saved Task configuration
• Run Task and get dataset items - fire and return results inline
• Scrape single URL - curated, 2-field wrapper over a content scraper (see section 10)
• Set key-value store record - write a file/string to a KV store

Searches (reads)

• Get last Actor/Task run - the most recent matching run
• Get run - by run ID
• List runs - filterable by status
• Get dataset items - by dataset ID, with offset/limit and optional fields/omit
• Get key-value store record - by store ID + record key

Hidden selectors (back dropdowns, not user steps)

• List Actors (recently used source + Store source)
• List Tasks

If the host platform cannot represent every operation, prioritize in this order: run-finished trigger, run Actor + get dataset, get dataset items, scrape single URL, run Task + get dataset, get KV record, set KV record.

3. Resource -> operation organization

Organize the action node’s surface as resource -> operation, mirroring both the host platform’s UX convention and Apify’s domain model. A two-level router (resource, then operation) keeps adding an operation a copy-and-adapt task and co-locates each operation’s parameters and I/O. Users navigate Actor -> “Run actor”, Dataset -> “Get items”, etc. Avoid a flat list of dozens of operations.

4. Actor / Task selection UX

IDs (Actor, Task, run, dataset, KV store, record key) should use the host’s resource-locator property type with multiple modes: From list, By URL, By ID. Extract the ID from a pasted Console URL (e.g. https://console.apify.com/actors/<id>/input -> <id>).

Offer two selection sources in one selector:

• Recently used (default) - the user’s own Actors/Tasks, sorted by last run started. This is what most users want.
• Store - browse the public marketplace by search term, sorted by popularity.

A toggle with “recompute fields when changed” lets users switch the dropdown source without leaving the node. Choose a pagination limit that balances completeness against load time (the host UI must support searching within dropdown results).

5. Dynamic input translation

The highest-leverage UX piece. When a user selects an Actor, fetch the Actor’s build and translate its inputSchema into the host platform’s form fields dynamically. Hardcoding fields per Actor is unmaintainable across thousands of Actors; a translator supports any Actor’s input as a host form.

Handle the type/editor mappings that have no direct host equivalent, and degrade gracefully rather than crashing:

Two gotchas:

• On list fields, a default would pre-populate a non-removable first item and duplicate entries - move prefill values to placeholder.
• Mark the build selector so a build-tag change recomputes the generated fields, since input schemas can differ across builds.

6. Run UX: sync vs async, cost caps, bounded polling

Offer a sync/async toggle on run-start actions:

• Sync: block until the run finishes (document the host’s hard timeout so users know when to choose async).
• Async: return immediately with the run ID for long-running jobs.

Use the asynchronous REST flow under the hood (POST /runs with waitForFinish=0, then poll). Polling must be bounded - use the run’s own timeoutSecs + a grace buffer, with an absolute ceiling fallback (e.g. 24h). Poll at a fixed interval (e.g. 1s). “Run actor and get dataset” must additionally require status === 'SUCCEEDED' before fetching dataset items.

Every run-start action exposes a Maximum Cost per Run field (maxTotalChargeUsd), min: 0, default null. Send it as a query parameter only when non-null and > 0; 0/empty means no limit. Never let this be an Actor input field.

7. Output normalization into one enriched shape

Bare Actor run objects contain only storage IDs, not results - useless downstream. Enrich every run once, centrally, into a single shape returned by all run-producing and run-finding actions:

• The run’s OUTPUT from the default key-value store (and any user-selected store keys).
• datasetItems - up to a sensible cap (e.g. 100 items) from the default dataset.
• datasetItemsFileUrls - pre-built download URLs for JSON/CSV/XML/XLSX/HTML/RSS exports.
• detailsPageUrl - a deep link to the run in Apify Console.
• Strip fields useless to host users (meta, stats, options, userId, raw output, standby).

Downstream steps map fields by name regardless of which action produced the run, because every action yields the same shape. Describe this shape to the host UI with sampled output fields (sample ~10 items, merge their keys).

8. Guard against large datasets

Estimate the full download size before inlining: fetch one item, multiply by the requested item count with a safety margin (e.g. 1.2x), compare to a payload cap (e.g. 10 MB). If it exceeds the cap, do not attempt the inline download - return a warning item plus the dataset file URLs. When a trigger fetches more items than the cap allows, push a warning pointing users to the “Get Dataset Items” action with run.defaultDatasetId for full results. The host platform has payload/time limits the Apify API does not; always offer a file-download fallback.

9. Files and binary records

Map the host’s file-handling primitives (dehydration / stashing / signed URLs) to Apify’s binary KV records. For “Get KV record”:

• JSON/text under a size threshold (e.g. 20 MB) -> parsed and returned as fields the user can map directly.
• Binary or larger records -> a dehydrated pointer the host lazily fetches when a downstream step needs it.
• Records above a hard limit (e.g. 120 MB, with margin below the host’s 150 MB ceiling) -> rejected outright with an actionable message.

10. Run-finished trigger UX

Use the host’s hook type backed by Apify webhooks:

• Subscribe: create an Apify webhook scoped to an actorId or actorTaskId, with eventTypes from the user-selected terminal statuses. The webhook’s requestUrl is the host’s target URL. Make registration idempotent - derive a key from every field that distinguishes one registration from another (resource:id:sortedEvents:requestUrl) so re-activating a workflow does not create duplicate webhooks. Persist the created webhook ID so deactivation can delete it.
• Unsubscribe: delete the webhook by its stored ID.
• Perform: read the webhook payload and enrich it (section 7).
• Fallback list: fetch the 3 most recent matching runs so users see realistic test data when configuring the trigger without waiting for a real event.

Offer an event multi-select (SUCCEEDED / FAILED / TIMED-OUT / ABORTED) plus an “any” shortcut that expands to all four. Build the condition: { actorId } or { actorTaskId } from the watched resource. Pass the webhook payload through as workflow data - it already carries the run metadata.

11. Error UX

Centralize API error mapping in one place, not per action. Match each Apify error to the host’s error category:

• 5xx and 429 -> retryable (so backoff applies).
• token-not-found / auth errors -> an authentication error the user must fix.
• full-permission-actor-not-approved -> a non-retryable error with the validated approval URL in the message (approval is a manual Console action; retrying cannot help).
• 404 on KV record GET/HEAD -> return empty rather than throw (a missing key is a valid “no data” result).
• Other 4xx -> a plain error with the API’s actual message, or a generic fallback.

Use retryable errors only when retrying can actually succeed. Keep error codes (like EPERM) out of any message field the host replaces, so the real text survives.

12. Auth UX

For consumer-facing automation platforms, prefer OAuth2 with PKCE over an API-token field. Users authorize through a browser; no raw token typing. Disable auto-refresh only if the host cannot surface a refresh failure gracefully - otherwise leave it on. Validate the token with a GET /v2/users/me test call and populate the connection label with the username/email.

If the host is headless-only, fall back to an API-token credential with the same verify call on a “Verify” button.

13. Convenience operations: “Scrape single URL”

Beyond generic “run Actor”, ship a curated Scrape single URL action: a 2-field form (url, outputFormat) wrapping a content scraper with sensible defaults (maxCrawlDepth: 0, maxResults: 1). Validate the URL first (new URL() + protocol + hostname check) with an actionable error - bad input must never start a paid run. Return a lean, single-object output: { ...pageMetadata, [outputFormat]: content } - strip all content variants and re-add only the chosen one. This is ideal for LLM flows (a single object beats a dataset array) and lowers the barrier for non-power users. Point power users to the underlying Actor for advanced options.

Definition-of-done checklist

• Not done: Capability matrix from section 2 is fully surfaced (or trimmed with documented reason).
• Not done: Resource -> operation organization; no flat operation soup.
• Not done: Actor/Task selection offers recently-used + Store sources with URL/ID extraction.
• Not done: Actor input is translated dynamically from the build schema; degradation is graceful.
• Not done: Run actions expose sync/async toggle and a maxTotalChargeUsd field (null = unlimited).
• Not done: Polling is bounded; never while (true).
• Not done: All run-producing and run-finding actions return the one enriched run shape.
• Not done: Large-dataset guard with file-URL fallback is in place.
• Not done: Run-finished trigger is webhook-backed, idempotent, and has fallback sample data.
• Not done: Error mapping is centralized; approval URLs are validated; codes don’t clobber messages.
• Not done: OAuth2 PKCE is the default consumer auth path; token fallback has a verify call.
• Not done: A “Scrape single URL”-style convenience action exists with pre-run URL validation.
• Not done: x-apify-integration-platform header is sent on every outbound request; x-apify-integration-origin: apify-integration-development-skill included if built from this skill.
• Not done: Two test modes (mocked + live E2E) pass.