5.8 KiB
title: Hub triggers description: How Hub matches an inbound event to a trigger: events, filters, and the allowlist that gates every execution. nav: Triggers order: 66 category: Hub
Triggers
A trigger says which provider event can start a workflow. The Hub workflows page covers the steps, inputs, routing, prompts, and deadlines that run after a match.
.paseo/workflows/github-issue.yml:
name: triage-issue
on: github.issue_created
filters:
repo: acme/api
from_users: [alice]
max_runtime: 2h
steps:
- id: work
environment: dev
max_runtime: 90m
idle_timeout: 10m
agent: codex
prompt:
- text: Call hub.finish_execution when the step is complete.
- text: ${{ paseo.prompt }}
Field-by-field detail is in the configuration reference.
Events
on |
Fires when |
|---|---|
github.issue_created |
An issue is opened. |
github.pull_request_created |
A pull request is opened. |
github.issue_comment_created |
A comment is created on an issue. |
github.pull_request_comment_created |
A conversation comment is created on a pull request. |
github.issue_label_added |
A label is added to an issue. |
github.pull_request_label_added |
A label is added to a pull request. |
slack.mention |
The bot is mentioned in a channel. |
discord.mention |
The bot is mentioned in a guild. |
manual.run |
A run started from the API. |
New GitHub workflows should use a semantic event. The five legacy events remain compatible: github.issues, github.issue_comment, github.pull_request_review, github.pull_request_review_comment, and github.push. See GitHub triggers for complete workflows and when to use each event.
Each provider page documents its events and the data they expose:
Filters
filters is required, and from_users must be present and non-empty. A trigger without it is rejected at validation.
The allowlist is what keeps a stranger's comment on a public issue from starting an agent on your machine. There is no default, because a safe default differs per repository.
An allowlist is one layer of defense. It does not make a permitted account trustworthy after compromise or make prompt injection harmless. See Hub security before choosing the daemon, working directory, provider policy, and outputs for an external trigger.
| Filter | Applies to | Matches |
|---|---|---|
from_users |
all | GitHub: login. Slack and Discord: user id, not display name |
repo |
GitHub | owner/name |
workspace |
Slack | Team id, T01234567 |
guild |
Discord | Guild id |
channels |
Slack, Discord | Channel ids |
contains |
all | GitHub substring; Slack and Discord invocation prefix |
pattern |
all | Invocation prefix |
connection |
all | A connection slug, when the organization has several |
label |
GitHub label-added events | The label added by this delivery, case-insensitively |
labels |
GitHub | Every listed current issue or pull-request label, case-insensitively |
All conditions must pass. There is no any mode.
Which connection an event comes from
repo, workspace, and guild are resolved to immutable ids when the configuration activates, along with the connection that owns them. Naming a resource the organization has no connection for fails activation, so you find out on push rather than when someone comments.
Omit the resource filter and the trigger listens to every connection of that provider in the organization. To pin it to one:
filters:
connection: acme-github
from_users: [alice]
See How Hub works for what activation compiles.
When two triggers match
Both run. Triggers are not ordered and do not shadow each other, in one configuration or across projects.
Replying
Put allow_outputs on the step that should reply. The reply capabilities are slack.reply and discord.reply.
- Set
maxwhen a step needs more than one update. - Set
required: truewhen the step must emit at least one reply before it can finish. A required type must be registered and available for the execution context.
GitHub has no reply capability; a step with a github block comments through gh instead. The output capability reference has the contract.
The declaration grants the hub.reply tool; the prompt has to tell the agent to call it. See Tell the agent which tool to call.