Airflow Hitl

Airflow Human-in-the-Loop Operators

Pause a DAG until a human responds via the Airflow UI or REST API. HITL operators are deferrable — they release their worker slot while waiting.

Requires Airflow 3.1+ (af config version).

>

UI location: Browse → Required Actions. Respond from the task instance page's Required Actions tab.

>

Cross-references: airflow-ai for AI/LLM task decorators; airflow for registry and API discovery commands used below.

---

Step 1 — Pick the capability you need

This is the only place class names are hardcoded. The provider adds, renames, and removes params across releases — do not copy parameter lists from memory. Fetch the current signature before writing code.

---

Step 2 — Discover the current signatures from the Airflow Registry

Before writing HITL code, run these to see the live roster and constructor params (see the airflow skill for the full af registry reference):

# Every HITL-related module in the standard provider
af registry modules standard \
  | jq '.modules[] | select(.import_path | test("\\.hitl\\.")) | {name, type, import_path, short_description, docs_url}'

# Constructor signatures: name, type, default, required, description
af registry parameters standard \
  | jq '.classes | to_entries[] | select(.key | test("\\.hitl\\.")) | {fqn: .key, parameters: .value.parameters}'

# Pin to the exact installed provider version
af config providers \
  | jq '.providers[] | select(.package_name == "apache-airflow-providers-standard") | .version'
# then: af registry parameters standard --version <VERSION>

If the registry shows a param that this skill does not mention, prefer the registry. If the registry shows a class that is not in Step 1, treat it as additive — the decision table above may be stale.

---

Step 3 — Canonical example (approval gate)

Starting point for any HITL task. Adapt by swapping the class name and params per Step 2.

from airflow.providers.standard.operators.hitl import ApprovalOperator
from airflow.sdk import dag, task, chain, Param
from pendulum import datetime

@dag(start_date=datetime(2025, 1, 1), schedule="@daily")
def approval_example():
    @task
    def prepare():
        return "Review quarterly report"

    approval = ApprovalOperator(
        task_id="approve_report",
        subject="Report Approval",
        body="{{ ti.xcom_pull(task_ids='prepare') }}",
        defaults="Approve",              # Auto-selected on timeout
        params={"comments": Param("", type="string")},
    )

    @task
    def after_approval(result):
        print(f"Decision: {result['chosen_options']}")

    chain(prepare(), approval)
    after_approval(approval.output)

approval_example()

For the other classes in Step 1, the shape is the same (task_id, subject, plus class-specific params). Verify each constructor through Step 2 — for example, HITLBranchOperator requires every option either to match a downstream task id directly or to be resolved via a mapping param surfaced in the registry.

---

Step 4 — Behavior contracts (stable across versions)

Timeout

• With defaults set: task succeeds on timeout, default option(s) selected.
• Without defaults: task fails on timeout.

Markdown + Jinja in body

body supports Markdown and is Jinja-templatable. Render XCom context directly:

body = """**Total Budget:** {{ ti.xcom_pull(task_ids='get_budget') }}

| Category | Amount |
|----------|--------|
| Marketing | $1M |
"""

Callbacks

All HITL operators accept the standard Airflow callback kwargs (on_success_callback, on_failure_callback, etc.).

Notifiers

HITL operators accept a notifiers list. Inside a notifier's notify(context) method, build a link to the pending task with HITLOperator.generate_link_to_ui_from_context(context, base_url=...).

Restricting who can respond

The parameter name and accepted identifier format depend on the active auth manager. Do not hardcode — check which one is active and which kwarg the current provider exposes:

af config show | jq '.auth_manager // .core.auth_manager'

Then look up the current kwarg in Step 2 (at the time of writing it is assigned_users, accepting identifiers in whatever format the active auth manager uses — Astro uses the Astro user ID, FabAuthManager uses email, SimpleAuthManager uses username).

---

Step 5 — Responding from external integrations

For Slack bots, custom apps, or scripts. Discover the live endpoint rather than hardcoding a path:

af api ls --filter hitl           # live endpoint list
af api spec \
  | jq '.paths | to_entries[] | select(.key | test("hitl"))'   # request/response schemas

The PATCH-to-respond pattern is stable; the exact path is discovered. Typical shape:

import os, requests

HOST = os.environ["AIRFLOW_HOST"]
TOKEN = os.environ["AIRFLOW_API_TOKEN"]
HEADERS = {"Authorization": f"Bearer {TOKEN}"}

# List pending — use the path from `af api ls --filter hitl`
requests.get(f"{HOST}/<path>", headers=HEADERS, params={"state": "pending"})

# Respond — same discovered path family, PATCH
requests.patch(
    f"{HOST}/<path>/{dag_id}/{run_id}/{task_id}",
    headers=HEADERS,
    json={"chosen_options": ["Approve"], "params_input": {"comments": "ok"}},
)

---

Step 6 — Safety checks

• [ ] Airflow version ≥ 3.1 (af config version).
• [ ] Constructor kwargs match the current registry output from Step 2 — no respondents-vs-assigned_users style drift.
• [ ] For branching: every option resolves to a downstream task id (directly or via the mapping kwarg from Step 2).
• [ ] Every value in defaults is also in options.
• [ ] execution_timeout set; defaults configured if timeout should succeed rather than fail.
• [ ] API token configured if external responders are part of the flow.

---

References

The upstream docs URL is surfaced per-module by the registry — do not hardcode:

af registry modules standard \
  | jq '.modules[] | select(.import_path | test("\\.hitl\\.")) | {name, docs_url}'

Related skills

• airflow — af registry, af api, af config command reference.
• airflow-ai — AI/LLM task decorators and GenAI patterns.
• authoring-dags — general DAG writing best practices.
• testing-dags — iterative test → debug → fix cycles.