find-artifacts

Files (1)

• SKILL.md

SKILL.md (first lines)

# Find Artifacts (artifact-first)

Use this skill when the operator asks:

- "what can I view?"
- "show artifacts for this run"
- "load this recording/video/report"
- "browse outputs from storage"

## Model

Artifact-first means:

1. Enumerate what exists in storage.
2. Group by discovered run prefix.
3. Classify each object with a render hint.
4. Never hide unknown formats; degrade to download.

No workflow registry, path allowlist, or known-type gate is required.

## Local NPA search

Use `npa studio search` without initializing a film project. Repeat `--bucket`
for explicitly selected buckets; omit it to enumerate credential-visible buckets.
Use repeated `--project` aliases or `--all-projects` for configured contexts.
`--discover-tenant` inventories the selected tenant and tests each discovered
bucket with the appropriate external credential context. It does not grant access.

```bash
npa studio search --project "<project-alias>" \
  --bucket "<artifact-bucket>" --prefix "runs/" --kind video \
  --read-metadata > artifact-search.json
```

Inspect `complete`, `discovery`, and every `sources` row. Exit 1 preserves results
but means listing or discovery was incomplete. Metadata failure is separate from
listing coverage. Tool/model metadata is declared provenance; verify the original
run manifest and file hash before attributing an output. `--query` searches object
keys, not media contents or model metadata. Keep source tuples and search output
private. No recurring poller belongs in this skill or the reusable framework.

Human guide: [Find artifacts](../../../docs/workbench/cookbooks/find-artifacts.md).

## Agent API flow

All calls are same-origin (`/api/...`) on the authenticated agent VM.

1. Discover runs:

```http
GET /api/artifacts/runs?prefix=&limit=100
```

Run discovery is durable-S3-first. A filtered search refreshes a cold, empty, or
stale bounded source index before applying `q`; the in-process cache is only an
accelerator. Follow the opaque `next_cursor` with the same query and source
selection. It traverses one immutable server-side snapshot and becomes stale
after an access refresh, backend restart/cache reset, or changed query/source
context; restart at page one in that case.

Do not interpret a partial access scope as a complete empty search. Observed
matches remain available, but incomplete responses keep `total_runs=null`,
`query_complete=false`, and `pagination_complete=false` and include source
errors for the unsearched scope. An absent exact id is a trustworthy 404 only
after complete effective-source discovery.

2. List artifacts for one run:

```http
GET /api/artifacts/run/{run_id}
```

Exact lookup searches every authorized effective source, including configured
direct run-parent sources. A configured source is the exact
`(project_id, bucket, resolved_prefix)` tuple where `resolved_prefix` is the
directory immediately above run ids. If multiple tuples contain the same run
id, preserve the `409 ambiguous_run_id` result and have the caller select one;
never pick an arbitrary or apparently more complete source.

3. Load one explicit artifact:

```http
POST /api/sim-viz/load-artifact
{
  "run_id": "run-prefix",
  "run_ref": "<server-issued opaque run reference>",
  "project_id": "<selected project>",
  "resource_bucket": "<selected bucket>",
  "resolved_prefix": "<selected prefix, possibly empty>",
  "source_selected": true,
  "key": "path/to/run/object"
}
```

Preserve this exact source tuple from list response to card, thumbnail, preview,
download, and load action. `s3_uri` is provenance only: URI-only or otherwise
unscoped media requests are rejected with a stable 400 response because bucket
membership alone is not run-level authorization. Follow every artifact-page
cursor before computing preference; the shipped UI does this automatically.

4. Confirm loaded state:

```http
GET /api/sim-viz/status
```

Check `artifact_key`, `artifact_render`, `artifact_uri`, and `rerun_ready`.

## Render hints

- `rerun` (`.rrd`)
- `video` (`.mp4`, `.webm`, `.mov`)
- `image` (`.png`, `.jpg`, `.jpeg`, `.gif`, `.webp`)
- `json`
- `text`
- `download` (fallback for unknown/new types)

Unknown types must stay visible/selectable.

From nebius/nebius-physical-ai, Apache-2.0.

More from nebius/nebius-physical-ai