name: teraslice-job-builder description: >- Author, review, lint, and validate a Teraslice job.json — the config that defines a job's lifecycle, workers, assets, and its pipeline of operations (one reader, zero-or-more processors, one sender) plus reusable apis. Use when a user wants to build a new Teraslice job, fix or review an existing job.json, understand a job config error (e.g. _api_name, asset/op name collisions, missing _op), or pick the right operations/asset bundles for a pipeline. Does not deploy the job — it hands off a validated job.json and the teraslice-cli command to register it. metadata: version: "1.0.1"

Teraslice job builder

A Teraslice job is a JSON config describing a data pipeline: a lifecycle, some workers, the assets (processor bundles) it needs, and an ordered operations array — one reader first, zero or more processors in the middle, and typically a sender last — plus any reusable apis.

This skill helps you author, lint, and validate a job.json. It does not register or start the job against a cluster; when the config is ready it hands you the teraslice-cli command to run yourself.

Documentation is fetched live — don't rely on stale copies

This skill's references/ files are pointers, not full copies. They carry a small cheat-sheet of stable facts, but for anything version-specific, any full field list, or any operation config you're unsure about, `WebFetch` the live doc URL — it's always current with the published docs. Prefer fetching over guessing.

External asset bundles (separate sites — fetch these for op-specific fields):

Workflow

1. Gather requirements. Ask the user:

• Lifecycle: once (batch — runs to completion then exits) or

persistent (streaming — runs until stopped)?

• Data source (→ reader) and destination (→ sender).
• Transforms needed in between (→ processors).
• Rough volume / worker count.
• Which asset bundles their cluster actually has installed — this

constrains the operations available. If unknown, note it; the job can't use an op from an uninstalled asset.

• Which connectors/connections the cluster has configured — reader and

sender ops reference a connection (e.g. "connection": "default") that must exist in the cluster's terafoundation.yaml. If the cluster is reachable, list them with curl '<cluster>:<port>/v1/cluster/connectors' (?type=<type> to filter, ?groupBy=type for the grouped view) instead of guessing. If unknown, note it; a reader/sender pointed at a missing connection will fail at registration. Note any connection tagged `is_state_cluster` / `is_asset_store` — see [Connection guardrails](#connection-guardrails).

1. Pick operations. Reader first, processors in the middle, sender last.

Consult references/operations.md and references/builtin-operations.md; for external bundle ops, WebFetch the bundle's site (table above). references/common-assets.md is the cheat sheet for what lives where.

1. Assemble the job.json. Start from the closest skeleton in

assets/templates/ (batch-once vs persistent-streaming vs custom-asset) rather than from scratch. Fill fields per references/job-schema.md (fetch the configuration page for the authoritative list).

1. Lint early and often. After every edit, run:

`` node scripts/lint-job.mjs path/to/job.json `` Zero-dependency structural check — safe to run anywhere. Fix what it flags before moving on.

1. Deep-validate before handoff. When the user is ready to test/register:

`` node scripts/deep-validate-job.mjs path/to/job.json ` Runs Teraslice's real top-level job-schema validation (validateJobConfig, the same check JobValidator uses for the job config) **if** @terascope/job-components` is resolvable (inside a teraslice checkout, or installed as a dep). It does not do full per-operation asset validation — that needs each op's asset on disk. If job-components isn't resolvable it reports clearly what it skipped — don't claim the job is fully validated when only the lint ran.

1. Handoff — do NOT deploy. Tell the user the command to run themselves:

`` teraslice-cli tjm register <cluster-alias> path/to/job.json # add --start to register and start in one step `` This skill has no cluster credentials and does not register jobs.

Connection guardrails

/v1/cluster/connectors tags connections with is_state_cluster: true (the Opensearch/Elasticsearch connection where Teraslice stores its own job state, analytics and logs) or is_asset_store: true (the connection used for asset storage). If a reader — and especially a sender — targets a connection with either flag set to `true`, STOP and warn the user before continuing. Writing job data into Teraslice's own state or asset store can corrupt or pollute the cluster's internal storage. Surface this every time it happens, get explicit confirmation that they really intend to read from / write to that connection, and recommend a separate data connection instead. Only proceed once they've confirmed.

Warn even when it's unavoidable. Dev environments commonly have a single connection of a given type — one Opensearch/Elasticsearch connection (which therefore must be the state cluster) or one S3 connection (which must be the asset store) — so any ES or S3 reader/sender will land on it. Warn the user anyway, even though it's obvious and there's no alternative connection. Never silently skip the warning just because the flagged connection is the only option; still make sure they know they're sharing the state cluster or asset store and have them confirm.

Reference index

Limitations

• External asset bundle details live in other repos and change on their

own release cadence. Treat any version-specific field from common-assets.md as "verify against the bundle's own docs" — fetch the live site rather than trusting a cached field list.

• Deep validation is strongest on top-level job fields. Per-operation

schema validation needs that op's asset resolvable locally, so op-specific fields from bundles the user hasn't installed won't be fully checked. The real final check is registering against an actual cluster.

• This skill authors and validates only — it does not register, start,

stop, or deploy jobs.