Skip to main content
Cron, not events. Org-wide, not per-repo. One configuration, scans everything.
JacobPEvans/claude-code-routines ships seven prompt files that run on schedule inside Anthropic’s cloud-hosted Claude Code sandbox. They discover repos under $GH_OWNER via gh search, do their work by calling the gh command-line tool, and report results to Slack via the official Slack MCP connector. There is no per-repo wiring. Drop a repo into the org and the routines pick it up on the next run.

​
The seven routines

​
How the Issue Solver picks work

The most ambitious routine. Six phases, all running in the same sandbox session:
1

Discover

Shell-only. gh search issues across $GH_OWNER for the last 90 days, open + unassigned. Scored via jq weights: bug=+50, good-first-issue=+40, type:security=-40, type:breaking=-40, recent-activity bonus.
2

Triage

Claude Sonnet classifies the top 5 by solvability + complexity (trivial/small/medium/large).
3

Investigate

A read-only subagent reads the affected code and locates the exact lines that need changes.
4

Implement

Pre-flight secret scan (grep -P against known patterns). Creates a branch via gh api repos/.../git/refs. Pushes file changes via Contents API. The App installation token auto-signs every commit.
5

Verify

Best-effort CI polling at 30-second intervals up to 5 minutes. Records the final ci_status (passed / failed / pending / none).
6

Submit

Opens a draft PR only: never --ready, never auto-merge. If the issue is too complex, it comments on the issue explaining what was tried and gives up cleanly.

​
How routines authenticate

Each routine runs in Anthropic’s cloud sandbox with the gh command-line tool pre-installed. Authentication is via environment variables set in the cloud routine configuration at claude.ai/code/routines: Slack output goes through the official Slack MCP connector. The full operator-setup walkthrough lives in docs/CLOUD_ROUTINES_AUTH.md in the source repo. The local Claude Code that develops and deploys these routines is configured declaratively via nix-claude-code. Nix version-pins the plugin set, MCP wiring, permissions, and hooks.

​
Current deploy gotcha

The native GitHub Actions deploy (.github/workflows/deploy-routines.yml) is turned off because anthropics/claude-code-action@v1 does not carry the organization UUID binding that the Routines API requires, so RemoteTrigger calls fail with Unable to resolve organization UUID. Workaround in use today: a local deploy-routine-changes skill (under .claude/skills/ in the routines repo) drives RemoteTrigger get/update/create calls during interactive Claude Code sessions. The interactive harness has the org binding the GHA runner lacks. The fallback is the /schedule update command-line flow if RemoteTrigger breaks in interactive mode too. When Anthropic fixes the OAuth token to carry org UUID, the workflow can be restored. Until then, deploys are manual + interactive.

​
What routines do NOT do

​
Where this fits relative to ai-workflows

The two systems are complementary. ai-workflows handles the per-event reaction, and claude-code-routines handles the org-wide sweep.

​
Where to go next

Source repo

The six prompt files, schedule manifest, deploy notes.

Cloud routine auth

Operator setup: env vars, Slack MCP, token strategy.

ai-workflows

The event-driven half of the same ecosystem.

Issue → PR pipeline

The per-repo cloud flow that runs in parallel.