Action Runner

The Action Runner is the bridge between what a 7B language model can reliably produce and what a production automation system actually needs. This page explains why it exists, how it works end-to-end, and what each of its 20 step types does.

Why it exists

Qwen 2.5 7B and Dolphin 3 8B cannot reliably produce the structured JSON function-call format that AnythingLLM’s native tool-calling pipeline expects. Under load, the formatting breaks. Tool calls get dropped. JSON is malformed.

Rather than fight the model’s limitations, GenieHelper routes around them. Instead of asking the model to call a tool directly, the system asks it to emit a plain-text ACTION tag:

[ACTION:scrape-profile:{"platform":"onlyfans","user_id":"123"}]

This is something 7B models do reliably — append a structured string to a response. The Action Runner intercepts that string, parses it, fetches the corresponding flow definition from the action_flows Directus collection, and executes the steps deterministically without involving the LLM again. Result: Reliable automation from an uncensored 7B model with no cloud dependency.

Architecture

Agent response stream
    │
    ├── Plain text (sent to user)
    │
    └── [ACTION:slug:{"params"}] ──► Action Runner interceptor
                                          │
                                          ▼
                                 Fetch flow from action_flows
                                 (Directus collection)
                                          │
                                          ▼
                                 Execute steps sequentially
                                 (20 step types, 120s timeout each)
                                          │
                                          ▼
                                 Return { success, results }
                                 to genieChat SSE stream

Source: server/utils/actionRunner/index.js · server/utils/actionRunner/stepExecutors.js The action bus integration lives in server/utils/actionBus/ and handles the BullMQ enqueue side.

ACTION tag format

The format is rigid and parseable by a simple regex — no LLM is involved in parsing:

[ACTION:<slug>:<json-params>]

Part	Description
`slug`	Unique identifier of a flow in the `action_flows` Directus collection
`json-params`	JSON object — the input parameters passed to the flow’s first step

Examples

[ACTION:scrape-profile:{"platform":"onlyfans","user_id":"123"}]

[ACTION:taxonomy-tag:{"content_id":"abc-456","text":"beach yoga selfie"}]

[ACTION:post-create:{"draft_id":"789","platform":"fansly","scheduled_at":"2026-03-21T14:00:00Z"}]

[ACTION:message-generate:{"fan_id":"fan-001","context":"subscriber renewal"}]

[ACTION:media-process:{"file_id":"media-999","operation":"watermark"}]

[ACTION:memory-recall:{"task":"scrape onlyfans stats","top_n":6}]

Flow definitions — the action_flows collection

Flows are stored as records in the action_flows Directus collection. Each record contains:

Field	Type	Description
`slug`	string (unique)	The identifier used in ACTION tags
`name`	string	Human-readable label for status messages
`active`	boolean	Disabled flows are rejected immediately
`steps`	JSON array	Ordered list of step definitions

Each step in the steps array has:

{
  "type": "directus_read",
  "result_key": "profile",
  "config": {
    "collection": "creator_profiles",
    "filter": { "id": { "_eq": "$input.creator_profile_id" }
    }
  }
}

The result_key value becomes available to subsequent steps as $result_key.field_name.

Seeded flows

The following flows are seeded in the action_flows collection at install time:

Slug	Purpose
`scout-analyze`	Analyse a creator’s platform presence and surface insights
`taxonomy-tag`	Run LLM-assisted taxonomy tagging on a piece of content
`post-create`	Create and optionally schedule a content post
`message-generate`	Draft a fan message using the creator’s persona
`memory-recall`	Activate relevant skills via JIT hydration
`media-process`	Enqueue a media job (watermark, clip, thumbnail)

The 20 step types

Step executors live in server/utils/actionRunner/stepExecutors.js. Each executor receives a config object (with interpolated variables) and an AbortSignal for the 120-second timeout.

directus_read — fetch items from Directus

Reads one or many items from a Directus collection.

{
  "type": "directus_read",
  "config": {
    "collection": "creator_profiles",
    "filter": { "id": { "_eq": "$input.creator_id" } },
    "fields": "id,display_name,platform_connections",
    "limit": 1
  }
}

Supports both single-item (id param) and list queries with filter, sort, fields, and limit.

directus_write — create a new item

Creates a new record in a Directus collection.

{
  "type": "directus_write",
  "config": {
    "collection": "media_jobs",
    "item": {
      "operation": "apply_watermark",
      "file_id": "$input.file_id",
      "status": "queued"
    }
  }
}

directus_update — PATCH an existing item

Updates a single item by ID.

{
  "type": "directus_update",
  "config": {
    "collection": "media_jobs",
    "id": "$job.id",
    "item": { "status": "processing" }
  }
}

directus_delete — remove an item

Deletes a single item from a collection.

{
  "type": "directus_delete",
  "config": {
    "collection": "hitl_sessions",
    "id": "$input.session_id"
  }
}

directus_search — full-text search

Full-text search across a collection with optional field projection and limit.

{
  "type": "directus_search",
  "config": {
    "collection": "fan_profiles",
    "query": "$input.query",
    "fields": "id,display_name,platform",
    "limit": 10
  }
}

directus_trigger — fire a Directus Flow by UUID

Triggers a Directus webhook flow. Note: the Directus request operation is known-broken (returns {}). This step triggers flows via their webhook endpoint instead.

{
  "type": "directus_trigger",
  "config": {
    "flow_uuid": "7506f825-...",
    "payload": { "content_id": "$input.content_id" }
  }
}

Known flow UUIDs: 591a8845 (memory-sync) · 7506f825 (taxonomy-tag) · 50a4dc14 (post-create) · a4f083fa (message-generate).

stagehand_start — launch a browser session

Starts a headless Playwright session via the Stagehand server and returns a sessionId.

{
  "type": "stagehand_start",
  "result_key": "browser",
  "config": {}
}

All subsequent Stagehand steps require $browser.sessionId.

stagehand_navigate — go to a URL

Navigates the browser session to a URL.

{
  "type": "stagehand_navigate",
  "config": {
    "sessionId": "$browser.sessionId",
    "url": "https://onlyfans.com/my-creator-handle"
  }
}

stagehand_act — natural language browser action

Executes a natural language action against the current page (click, type, scroll, select).

{
  "type": "stagehand_act",
  "config": {
    "sessionId": "$browser.sessionId",
    "action": "click the Stats tab"
  }
}

stagehand_extract — extract structured data

Extracts structured data from the current page using a natural language instruction and optional JSON schema.

{
  "type": "stagehand_extract",
  "result_key": "stats",
  "config": {
    "sessionId": "$browser.sessionId",
    "instruction": "Extract subscriber count, total earnings, and post count",
    "schema": {
      "subscribers": "number",
      "total_earnings": "string",
      "post_count": "number"
    }
  }
}

stagehand_close — end the browser session

Closes and cleans up a Stagehand browser session.

{
  "type": "stagehand_close",
  "config": { "sessionId": "$browser.sessionId" }
}

stagehand_cookie_login — authenticate using stored cookies

Decrypts the stored AES-256-GCM credentials for a creator profile, injects the cookies into a new browser session, and navigates to the target platform URL. Returns { sessionId, navigated: true }.

{
  "type": "stagehand_cookie_login",
  "result_key": "session",
  "config": {
    "creator_profile_id": "$input.creator_profile_id",
    "platform": "onlyfans",
    "target_url": "https://onlyfans.com/my/statistics"
  }
}

Used in scrape flows instead of stagehand_start + stagehand_navigate when authenticated sessions are required.

stagehand_extract_cached — extract with cache check

Like stagehand_extract but checks extractCache first. Returns the cached result if the (sessionId, instruction) pair was recently extracted, otherwise performs a live extraction and caches the result.

{
  "type": "stagehand_extract_cached",
  "result_key": "stats",
  "config": {
    "sessionId": "$session.sessionId",
    "instruction": "Extract subscriber count and total earnings",
    "schema": { "subscribers": "number", "earnings": "string" }
  }
}

ollama_chat — multi-turn conversation

Runs a multi-turn chat completion through Ollama. Use this instead of ollama_generate when conversation history matters.

{
  "type": "ollama_chat",
  "result_key": "reply",
  "config": {
    "model": "dolphin-mistral:7b",
    "messages": [
      { "role": "system", "content": "Write in the creator's voice." },
      { "role": "user", "content": "$input.fan_message" }
    ]
  }
}

taxonomy_terms_read — cached taxonomy lookup

Reads taxonomy terms for a given platform from taxonomyCache. Faster than directus_read on taxonomy_mapping for hot-path flows because it bypasses Directus and reads the in-process cache.

{
  "type": "taxonomy_terms_read",
  "result_key": "terms",
  "config": { "platform": "onlyfans" }
}

taxonomy_platform_maps_read — cached platform taxonomy maps

Reads platform-to-taxonomy mapping for a given platform from taxonomyCache.

{
  "type": "taxonomy_platform_maps_read",
  "result_key": "maps",
  "config": { "platform": "fansly" }
}

taxonomy_invalidate — bust the taxonomy cache

Invalidates the in-process taxonomy cache. Call this after a taxonomy graph rebuild or bulk ingest to force the next request to re-read from Directus.

{
  "type": "taxonomy_invalidate",
  "config": {}
}

ollama_generate — single-turn LLM completion

Runs a single-turn prompt through Ollama. Used for sub-tasks within a flow that need LLM output without a full agent session.

{
  "type": "ollama_generate",
  "result_key": "caption",
  "config": {
    "model": "dolphin-mistral:7b",
    "system": "Write in the creator's voice. Style: playful. No hashtags.",
    "prompt": "Write a caption for a beach selfie. Stats: $stats"
  }
}

bullmq_enqueue — push a job to BullMQ

Enqueues a job to the media-jobs BullMQ queue. The media-worker consumer picks it up and executes FFmpeg, ImageMagick, or Stagehand operations.

{
  "type": "bullmq_enqueue",
  "config": {
    "queue": "media-jobs",
    "job": {
      "type": "apply_watermark",
      "file_id": "$input.file_id",
      "creator_profile_id": "$input.creator_id"
    }
  }
}

The executor returns { queued: true, bullId, jobId }. Use media.job-status via MCP to poll completion.

http_request — generic HTTP call

Makes an arbitrary HTTP request. Used for webhooks, internal API calls, and third-party integrations not covered by the other step types.

{
  "type": "http_request",
  "config": {
    "url": "http://127.0.0.1:3001/api/some-internal-endpoint",
    "method": "POST",
    "body": { "user_id": "$context.user_id" }
  }
}

Variable interpolation

Step configs support variable substitution using $ prefixes. Interpolation is resolved at execution time, just before each step runs.

Variable syntax	Resolves to
`$input.field`	A field from the ACTION tag’s JSON params
`$steps[N].field`	The result of step N (zero-indexed)
`$result_key.field`	The named result of a previous step
`$context.field`	Extra context injected by genieChat (user info, workspace)

// Step 1 result_key = "profile"
// Step 2 can reference it as:
{ "creator_id": "$profile.id" }

End-to-end example: “scrape my OnlyFans stats”

A creator sends this message in the chat interface. Here is what happens:

Agent emits ACTION tag

The Dolphin 3 8B orchestrator model processes the request and appends an ACTION tag to its response:

I'll scrape your OnlyFans stats now.

[ACTION:scrape-profile:{"platform":"onlyfans","creator_profile_id":"cp-123"}]

genieChat intercepts the tag

The streaming response is scanned for the [ACTION:...] pattern. When found, the tag is extracted and passed to ActionFlowExecutor.run("scrape-profile", { platform: "onlyfans", creator_profile_id: "cp-123" }).

Flow definition fetched from Directus

The executor queries action_flows for slug = "scrape-profile". It retrieves a 5-step flow definition: read credentials → start browser → inject cookies → navigate → extract stats.

Steps execute sequentially

Each step runs with a 120-second timeout:

directus_read — fetches the platform_sessions record for this creator
stagehand_start — launches a headless browser, returns sessionId
stagehand_navigate — navigates to https://onlyfans.com/my-handle/stats
stagehand_extract — extracts subscriber count, earnings, post count as JSON
directus_write — writes extracted stats to platform_stats_snapshots

BullMQ job enqueued

A final bullmq_enqueue step creates a scrape_post_performance job so individual post metrics are fetched asynchronously without blocking the chat response.

SSE stream updates the UI

As each step completes, onStatus callbacks emit SSE events that update the TopCueRail mission label and open the live-scrape-feed panel in the Right Wing automatically.

Error handling

Each step has a 120-second hard timeout enforced via AbortController.
If any step fails, the executor returns { success: false, failedStep: N, error: "..." } immediately — subsequent steps do not run.
Timeout and failure messages are surfaced via onStatus to the SSE stream, so the creator sees a meaningful error rather than a stalled spinner.
Disabled flows (active: false) are rejected before any steps run.

The Action Runner authenticates to Directus using the admin credentials in DIRECTUS_EMAIL / DIRECTUS_PASSWORD. These are not the creator’s credentials. The user_id context injected by genieChat.js must come from the validated session — never from the LLM’s output.

AI System

Memory & Retrieval

Taxonomy

Why it exists

Architecture

ACTION tag format

Examples

Flow definitions — the action_flows collection

Seeded flows

The 20 step types

Variable interpolation

End-to-end example: “scrape my OnlyFans stats”

Error handling

Build docs developers (and LLMs) love

AI System

Memory & Retrieval

Taxonomy

​Why it exists

​Architecture

​ACTION tag format

​Examples

​Flow definitions — the action_flows collection

​Seeded flows

​The 20 step types

​Variable interpolation

​End-to-end example: “scrape my OnlyFans stats”

​Error handling

Build docs developers (and LLMs) love

Why it exists

Architecture

ACTION tag format

Examples

Flow definitions — the action_flows collection

Seeded flows

The 20 step types

Variable interpolation

End-to-end example: “scrape my OnlyFans stats”

Error handling