vitna.ai/docs/cli/witness/collect Machine view
Raw .md llms.txt
> Source: https://vitna.ai/docs/cli/witness/collect
> Raw Markdown: https://vitna.ai/docs/cli/witness/collect.md
> Corpus index: https://vitna.ai/llms.txt

# npm run witness:collect

One collector pass: fetch every enabled source, parse it, archive it, and write
what changed.

```bash
npm run witness:collect
npm run witness:collect -- --dry-run
npm run witness:collect -- --source anthropic-pricing
```

## Flags

| Flag | Meaning |
|---|---|
| `--dry-run` | Needs no database. Fetches and parses, writes nothing. |
| `--source <id>` | Read one source, enabled or not |

`--source` reading a disabled source is what pairs with
[`witness:seed --hold-new`](seed.md): a new adapter gets watched against
the live page before it is turned on for everyone.

## What one pass writes

A `fetch_run`, a `fetch` row per source with a final status, an
[observation](../../concepts/observation.md) per parsed figure, and whatever
[price periods](../../concepts/price-period.md) and
[market events](../../concepts/market-event.md) the comparison implies.

A priced row the registry does not hold becomes a
[candidate](../skus/candidates.md) rather than an observation.

## In-flight runs and freshness

A run's row is written when it finishes. While it is working, the source's last
completed run is still the most recent one, so a source can read as stale while
it is actively being read.

Check whether a run is live before acting on a stale reading. Re-enqueueing a
job that is eighty percent through a backfill makes it worse.

## The engine runs at the end

Trigger evaluation happens after the collector's writes, in the same pass, so a
price move and the watch it fires are one unit of work rather than two that can
disagree.
Human Machine Shift+M