[FEATURE] Native support for Agent Events — portable, scheduled AI workflows discoverable from a project directory

▎ Suggested labels: enhancement · area:scheduling · area:hooks

Preflight Checklist

• [x] I have searched existing requests and this feature hasn't been requested yet
• [x] This is a single feature request (not multiple features)

Problem Statement

## Summary

Claude Code recently shipped /loop and CronCreate for session-scoped recurring tasks. This is a great foundation —
we'd like to propose extending it with support for Agent Events: a portable, file-based format for defining
scheduled AI workflows that live inside a project repository and survive across sessions, machines, and agent
platforms.

## The Problem

Today, scheduled agent work in Claude Code is:

• Session-bound — /loop tasks vanish when the session ends
• Not portable — no standard format to share scheduled workflows across teams or repos
• Not composable — no way to bundle a scheduled task with the skills, references, and scripts it needs
• Not auditable at the project level — schedules exist outside version control

Proposed Solution

## The Proposal: Agent Events Format

An Agent Event is a simple folder inside an events/ directory:

``
events/
└── daily-standup/
├── EVENT.MD # schedule + instructions (required)
├── skills/ # bundled agent skills (optional)
├── references/ # markdown references, configs (optional)
└── scripts/ # helper scripts (optional)
``

The EVENT.MD file uses YAML frontmatter + Markdown instructions — a format Claude Code already handles natively:

```markdown
---
name: daily-standup
description: Compile and post daily team status
schedule: "0 9 MON-FRI"
enabled: true
---

## Instructions

Gather updates from the last 24 hours and post a standup summary to Slack #engineering.

• Check GitHub for merged PRs and open reviews
• Summarize blockers from open issues labeled blocker
• Tag team members with pending review requests

```

That's it. No YAML sprawl, no separate config files. Anyone can read and edit it.

## What Native Support Would Look Like

We are not asking for a complex orchestration engine. The minimal viable integration would be:

1. Discovery — on session start (or /events command), Claude Code scans the project's events/ directory
2. Registration — qualifying events with enabled: true are registered as durable CronCreate tasks (the

infrastructure already exists)

1. Execution — when a cron fires, Claude runs the EVENT.MD instructions with any bundled skills and references

loaded into context

1. Persistence — registered events survive restarts (using the existing durable task mechanism)

The format is deliberately minimal so this can be a thin layer on top of what Claude Code already does.

## Why This Matters

| Today | With Agent Events |
|---|---|
| /loop tasks die with the session | Events survive restarts — they live in the repo |
| Schedules are not in version control | events/ is committed alongside code |
| No standard format to share with teammates | Clone the repo, get the events |
| Skills and context must be re-specified each time | Bundled with the event definition |
| Hard to audit what ran and when | Full history, versionable |

## Relationship to Existing Features

This is not a replacement for /loop or CronCreate — it is a structured convention on top of them, analogous to how
CLAUDE.md is a convention on top of the system prompt. Just as Claude Code reads CLAUDE.md on startup to
understand the project, it would read events/ to understand what scheduled work the project expects.

Alternative Solutions

_No response_

Priority

Critical - Blocking my work

Feature Category

CLI commands and flags

Use Case Example

Real-World Use Cases in Our Codebase

We have been using this format to define:

• daily-standup — weekday 9 AM digest of GitHub + Slack activity
• system-health-check — every 15 min, monitor thresholds and open PagerDuty incidents
• code-review-reminder — every 4 hours, DM reviewers with stale PRs
• sprint-summary — Friday 4 PM, generate and send sprint report

Additional Context

_No response_