scheduled-execution

scheduled-execution

TL;DR

Three tiers, picked by cadence and trust requirements:

Tier	Mechanism	Use when…
Cloud	Claude Code Routines (April 2026+)	Daily / hourly recurring work; runs on Anthropic infrastructure; no local machine needed
Local persistent	Desktop tasks (local Claude Code)	Daily / hourly recurring work; needs filesystem or local-only resource access
Session-scoped	/loop slash command in a session	Short-lived polling within an active session ("check the build every 5 minutes")

Reality-check (the daily Orryx Group audit at D:\Orryx\reality-check\) is the canonical example of tier 2 / Local Persistent. The architecture brief's Wave 4 reality-check plugin formalises this pattern.

Tier 1 — Cloud (Routines)

Best when:

• Work doesn't need local filesystem (or filesystem access via MCP is sufficient)
• You want it to run regardless of whether your machine is on
• The work is well-bounded (no human input needed)

Example use: daily audit of public-facing endpoints; daily Slack summary; weekly competitor analysis.

How to set up: Use the /schedule skill (Claude Code plugin) or Anthropic Console Routines panel.

Cost: per-run API cost. Tracks in usage dashboard.

Tier 2 — Local persistent (Desktop tasks)

Best when:

• Work needs local filesystem (e.g., scanning D:\ for changes)
• Work needs local-only secrets (e.g., a local SSH key)
• You're OK with it skipping when your machine is off
• You want full visibility into what ran

Example use: reality-check daily audit (scans D:\ project dirs); local backup verification; nightly local lint pass.

How to set up:

# Windows Task Scheduler — example for reality-check
$action = New-ScheduledTaskAction -Execute "claude" -Argument "code /reality-check"
$trigger = New-ScheduledTaskTrigger -Daily -At 06:00
Register-ScheduledTask -TaskName "Orryx-RealityCheck" -Action $action -Trigger $trigger

For macOS/Linux: launchd or cron respectively.

Cost: free except the API cost per run.

Tier 3 — Session-scoped (/loop)

Best when:

• Polling for a state change ("wait for the build to finish")
• Short-lived investigation that needs periodic check-ins
• You're actively in a session and want to stay engaged

Example use: "every 5 minutes check the deploy status until it's green or failed", "poll the PR for review comments", "wait for the test suite to complete then summarise".

How to use:

/loop 5m /deploy-status

Slash command runs every 5 minutes within the current session. /loop without an interval lets the model self-pace.

Cost: per-invocation API cost. Session ends → loop ends.

Decision flow

Is this a one-time task? → Just do it.
Does it need to run when my machine is off? → Tier 1 (Cloud Routines)
Does it need local filesystem / local secrets? → Tier 2 (Desktop tasks)
Is this just polling within a session? → Tier 3 (/loop)

Reality-check as worked example

D:\Orryx\reality-check\ runs daily at 0913 (per the audit's Phase 1 finding). It:

1. Scans 7 active Orryx repos for changes since last baseline
2. Updates baselines/<repo>.json (fingerprint-keyed)
3. Writes reports/daily-reality-check.md (rolling summary)
4. Archives previous runs to reports/archive/<run-id>.md

Tier 2 (Desktop tasks) is correct because:

• Needs filesystem access to D:\
• OK to skip when machine is off
• Has a defined work unit (one run per day)

Once delivery/reality-check (Wave 4 plugin) ships, this should consume the audit's _entity-map.json for per-entity rollups.

Production cadences (per repo)

Cadence	Use cases
Daily	Reality-check; security scan; dependency audit; backlog grooming
Hourly	Health checks; cost-spike detection; lint of in-flight branches
Weekly	Sprint planning prep (per methodology/spec-driven-delivery); cross-repo drift audit
Per-PR	Code review; cross-model verification; quality gates

Anti-patterns

• Tier-mismatch. Running a daily filesystem audit via cloud Routines that can't see your D:\ — won't work. Tier 1 for cloud-accessible data only.
• Long-running /loop in unattended session. Tier 3 is for active sessions. For unattended, use Tier 2.
• Cron-style without timezone awareness. Australian timezone matters for Triora's regulatory reporting. Document the timezone with the schedule.
• No monitoring on tier 2 tasks. If a Desktop task silently stops running, you may not notice for weeks. Add a "last successful run" heartbeat (reality-check does this).

See also

• delivery/reality-check (Wave 4) — the canonical Tier 2 example
• delivery/credential-audit (Wave 4) — could run at any tier
• delivery/security-audit (Wave 4) — typically daily Tier 2
• Anthropic Routines docs (April 2026+)
• Windows Task Scheduler / launchd / cron docs
• D:\Orryx\reality-check\ — local reference implementation