Dataset worked examples

Prompt-first procedure: Describe the outcome you want in your agent conversation. The agent should select and load the appropriate AIWG assets, explain material changes, request any needed approval, and report verification evidence. Exact commands and flags appear only in the CLI reference.

These examples use the shipped lifecycle surface. Confirm the installed surface with the agent-owned dataset operation.

Local file or directory

Ask: “Make `./records/` searchable, cite the file and revision for every result, keep processing local, and preview no more than 20 records.” The source assessment should reject traversal outside the declared root, enumerate symlink policy, propose a schema, and separate canonical files from the regenerable index.

CSV and JSONL

For `events.csv`, bind header names, delimiter, encoding, null rules, and a stable record identity. For `events.jsonl`, bind one JSON value per line, maximum line size, malformed-record policy, and the record schema. A preview must report rejected rows without silently changing their shape.

Authenticated API

Use an opaque locator such as `env:DATASET_API_TOKEN`; never place the token in the source definition, plan, logs, preview, fixture, or receipt. Declare the host allowlist, pagination and rate limits, retry policy, TLS requirements, and checkpoint semantics. This guide deliberately provides no live-network command or credential value.

Incremental update

Bind the adapter version, source schema, plan digest, prior committed receipt, and opaque cursor. Changed schema or plan identity requires a new reviewed plan. Missing records are not deletions unless the approved source contract defines tombstones.

Offline operation

Require `network: offline` and a local adapter. A warm, verified cache may answer a query for its bound revision. A cold, stale, corrupt, wrong-revision, or unverifiable cache must not be presented as current. See offline states.

Backend migration or rebuild

Preserve canonical sources and receipts, inventory derived artifacts, preview the mapping and loss report, build the replacement beside the old backend, and verify equivalent bounded queries before switching. Roll back by restoring the old selection; do not copy a cache and call it canonical data.