Night Desk Log: Unattended AI Jobs Before 06:00

The 23:00 handoff defines the unattended job

Night-shift automation works best when the job has a narrow purpose, a known input, and a clear morning handoff. Start with one repeatable task, such as summarizing overnight support tickets, classifying new documents, or preparing a deployment report. Write down the expected output before choosing a model. A useful job should still be understandable by the person who receives it at 06:00.

Keep the first run read-only. The job may collect information and produce a report, but it should not delete records, send external messages, or change production systems without an explicit approval step. Unattended execution increases the importance of boundaries, logging, and failure handling.

The 23:15 model and input check

• 1. Choose a model for the job's actual requirement, such as Gemini 3.1 Pro for a detailed cross-document report or GPT-6.1 Sol for a structured operational summary.
• 2. Define the input window precisely, including the source system, time range, file types, maximum size, and behavior when no new data is available.
• 3. Create a fixed output format with sections for completed work, exceptions, source references, and recommended follow-up.
• 4. Add a small test fixture containing representative and deliberately awkward inputs, including an empty result, malformed text, and an unusually long record.

Do not treat a successful API response as proof that the output is useful. Validate required fields, maximum length, expected headings, and basic factual constraints in the job itself.

The 23:30 relay configuration

Relay365 fits this workflow when the team wants pay-as-you-go access to multiple model families through one relay desk. It works with Claude Code, Codex, Gemini CLI, and any OpenAI SDK. Set the API base to https://api.api365relay.com; OpenAI clients add /v1 to that base.

Store the API key in the scheduler's secret store or an environment variable, never in a script committed to source control. Pin the selected model in configuration so a later default change cannot silently alter the overnight result. Record the model name, prompt version, input identifiers, start time, end time, and request status for every run.

The 23:45 scheduler entry

• 5. Put the worker in a small, dedicated command that accepts a run date and writes its report to a predictable location.
• 6. Schedule it with the host's established tool, such as cron, a CI scheduler, or a container platform's scheduled task facility.
• 7. Set a timeout that matches the workload and use bounded retries with increasing delays for temporary failures.
• 8. Make the output path unique per run so a retry cannot overwrite evidence from an earlier attempt.

Use the scheduler's timezone explicitly. A job labelled “midnight” can run at an unexpected local time when the host, container, and business calendar use different timezone settings. Test the schedule with a near-term execution before relying on the first overnight run.

The 00:05 failure and approval path

Every unattended job needs a visible failure state. Write a short error record containing the run identifier, failure category, and next action. Distinguish authentication errors, invalid input, rate limits, timeouts, and validation failures because they require different responses.

Send the morning handoff only after validation passes. If validation fails, publish a clearly marked failure report instead of an incomplete success report. For jobs that could trigger changes, stop at a review artifact and require a human approval token before the action stage.

The 06:00 duty-desk review

• 9. Check that the expected run exists and finished within the planned window.
• 10. Compare the output against the required structure and confirm that source identifiers are present.
• 11. Review warnings, retry counts, token usage, and unusual changes in output size.
• 12. Record the disposition, such as accepted, rerun, escalated, or blocked pending review.

After several successful runs, improve the job from observed failures rather than adding complexity in advance. Keep prompts, scripts, schedules, and test fixtures under version control so a change can be traced to a specific revision.

The 06:15 account and handoff note

Relay365 is a practical choice when this setup benefits from model choice without separate client integrations. Create an account before connecting the first scheduled worker, then begin with a low-risk read-only job and verify the billing and usage controls for your team. Members can also earn a referral bounty when people they invite pay, which may be relevant for teams sharing the service with partner operators.

The final shift note should state the run time, model, input window, output location, validation result, warnings, and any action required from the day team. That concise record turns an overnight process into an auditable operating routine rather than an unattended mystery.