# Nightly Pipeline Scheduler

Auto-starts batch pipelines when inference is available.

## What It Does

1. Checks inference provider health (OpenRouter, Ollama, RunPod)
2. Checks if it's off-peak hours (configurable, default: after 6PM)
3. Checks interactive session load (don't fight with live users)
4. Checks daily token budget (configurable limit)
5. Starts the highest-priority incomplete pipeline

## Pipeline Priority Order

| Priority | Pipeline | Deps | Max Tokens |
|----------|----------|------|------------|
| 1 | playground-factory | none | 100,000 |
| 2 | training-factory | none | 150,000 |
| 3 | knowledge-mine | training-factory running | 80,000 |
| 4 | adversary | knowledge-mine running | 50,000 |
| 5 | codebase-genome | none | 120,000 |

## Usage

```bash
# Normal run (used by cron)
./scripts/nightly-pipeline-scheduler.sh

# Dry run (show what would start)
./scripts/nightly-pipeline-scheduler.sh --dry-run

# Status report
./scripts/nightly-pipeline-scheduler.sh --status

# Force start during peak hours
./scripts/nightly-pipeline-scheduler.sh --force
```

## Configuration

Set via environment variables:
- `PIPELINE_TOKEN_LIMIT`: Daily token budget (default: 500,000)
- `PIPELINE_PEAK_START`: Peak hours start (default: 9)
- `PIPELINE_PEAK_END`: Peak hours end (default: 18)
- `HERMES_HOME`: Hermes home directory (default: ~/.hermes)

## Cron

Runs every 30 minutes. Off-peak only (unless --force).
See `cron/pipeline-scheduler.yml`.