vs-app-dataset-bind

name	vs-app-dataset-bind
description	Binds a dataset to an application with reviewed field config inference. Invoke when using `vs app dataset bind` and searchable/filter/suggest fields must be inferred and confirmed first.
category	app
applies_to	codex, agents, external-agent
requires_cli	>=0.1.0
keywords	app dataset bind, bind dataset, field config inference, index fields, filter fields, suggest fields
commands	app dataset bind, app dataset-config get, app dataset-config list, dataset schema get, dataset get

Viking App Dataset Bind

When to Use

Use this skill when the user wants to run vs app dataset bind and the bind step is responsible for inferring and confirming application-facing field settings instead of blindly reusing or omitting them.

Preconditions

• an application-id and a dataset-id are available
• you can access enough representative dataset context to infer bind-time field config
• prefer an existing plan directory or raw source file when available, because the inference prompts need Dataset_Description, Data, and Meaning
• schema and dataset-side field attributes have already been shown to the user and explicitly reconfirmed before bind-time inference starts
• do not assume the dataset-creation field-config.json already contains bind-time fields such as IndexFields, FilterFields, SuggestFields, ImageIndexFields, or VideoIndexFields
• if pulled dataset metadata already contains IndexFields, FilterFields, SuggestFields, ImageIndexFields, or VideoIndexFields, treat them as legacy/non-standard unless a human reviewer explicitly validates them

Commands

• app dataset bind: bind the dataset to the application with an explicit --field-config
• app dataset-config get / app dataset-config list: inspect current bound dataset config before changing it
• dataset schema get / dataset get: inspect the dataset schema and metadata needed for field inference

Workflow

1. Start by inspecting the current app and dataset context: existing dataset config, dataset schema, and any available source plan or raw sample data
2. Build the three prompt inputs required by the inference flow:
   • Dataset_Description: prefer an existing dataset description or plan output; if missing, derive it from the same source context used for binding
   • Data: collect up to 10 representative JSON/JSONL records
   • Meaning: prepare field meanings in the same tree shape expected by prompt inference
3. Reuse /Users/bytedance/go/src/SearchCLI/skills/schema.prompt and apply these prompt templates:
   • filter_infer_prompt to infer FilterFields
   • suggest_infer_prompt to infer SuggestFields
   • index_infer_prompt to infer IndexFields
4. Assemble a proposed field config JSON for bind:
   • keep the dataset-side FieldDescMap if it already exists
   • ignore pulled IndexFields from dataset metadata when preparing bind-time config
   • set IndexFields from the search output of index_infer_prompt
   • ignore pulled FilterFields from dataset metadata when preparing bind-time config
   • re-infer FilterFields from filter_infer_prompt, grounded in the field descriptions from FieldDescMap
   • ignore pulled SuggestFields from dataset metadata when preparing bind-time config
   • set SuggestFields from the suggest output of suggest_infer_prompt
   • ignore pulled ImageIndexFields and VideoIndexFields from dataset metadata when preparing bind-time config
   • infer ImageIndexFields from schema, meanings, and sample values by selecting true image URL or image asset fields
   • infer VideoIndexFields from schema, meanings, and sample values by selecting true video URL or media asset fields
5. Show the proposed IndexFields, FilterFields, SuggestFields, , and to the user and explicitly confirm them before binding

Video Dataset Field Constraints (DefaultFieldStrategy)

Video datasets (Dataset.Type = 3, produced with --type video in the onboarding skill) carry HARD server-side constraints on the bind-time field groups. Violating them causes BindAppDataset / CreateDataset to fail with:

Error.Code    = MissingParameter.DefaultFieldStrategy
Error.Message = The default field strategy is missing.

These rules are enforced by checkFieldConstraint on the server and cannot be bypassed by inference, prompt tuning, or user preference. Any bind-time proposal that violates them must be fixed before binding.

Field	IndexFields	FilterFields	SuggestFields
video_url	Must	Forbidden	Forbidden
content_id	Forbidden	Must	Forbidden
content_type	Forbidden	Must	Forbidden
parent_content_id	Forbidden	Must	Forbidden
sequence_index	Forbidden	Must	Forbidden

• Must → the field MUST appear in that group.
• Forbidden → the field MUST NOT appear in that group.
• Other fields (title, director, actors, genres, duration, language, alias, …) carry no server constraint; choose them based on the business goal and the inferred output.
• ImageIndexFields and VideoIndexFields are selected independently from the actual image / video asset fields in the schema. video_url is commonly added to VideoIndexFields as well; that is independent of the table above.

Before calling vs app dataset bind, the agent MUST verify every row of the table against the current proposal. If the proposal violates any row:

1. fix the proposal (move / add / remove the constrained fields),
2. re-render the affected per-group table,
3. ask the user to reconfirm through the Stage B dialog,
4. only then run vs app dataset bind.

See the authoritative reference at vs-item-onboarding/references/video-field-constraints.md.

Customer Environment Principle

• In customer environments, assume repository source code is unavailable.
• Execute tasks using only the installed skills, the packaged vs CLI surface (--help, command output, and observed runtime behavior), and explicit user-provided information.
• Do not rely on reading local repository source files, generated repo snapshots, or implementation details to decide runtime actions.
• If the installed CLI behavior conflicts with a skill, trust the installed CLI behavior first.
• If the skills and the packaged CLI still do not provide enough information to proceed safely, stop and ask the user instead of searching source code.

Constraints

• Do not run app dataset bind with a guessed or empty bind-time field-config when searchable/filter/suggest/image/video fields are part of the user request
• Do not start bind-time inference until schema and dataset-side field attributes have been explicitly reconfirmed with the user
• Do not skip user confirmation for IndexFields, FilterFields, SuggestFields, ImageIndexFields, or VideoIndexFields
• When the bind is driven by an agent chat, confirmation MUST use per-group tables plus an interactive question dialog as specified in vs-item-onboarding/references/agent-confirmation-ux.md; do not downgrade to a single plain-text summary, a free-form yes/no chat reply, or the CLI terminal y/N
• Do not auto-confirm a CLI summary step on the user's behalf. The agent-side table + dialog review remains the source of truth
• Reuse the prompt contracts in schema.prompt; do not invent alternate output keys for these inference steps
• Treat dataset creation and dataset binding as separate stages: FieldDescMap belongs to dataset creation, while IndexFields, FilterFields, SuggestFields, ImageIndexFields, and VideoIndexFields are inferred at bind time
• Treat pulled dataset-side IndexFields, FilterFields, SuggestFields, ImageIndexFields, and VideoIndexFields as untrusted historical data by default; use FieldDescMap plus representative samples to re-infer the bind-time field groups
• Treat prompt output as a proposal, not ground truth; remove noisy fields when the sampled data or business goal contradicts the prompt result
• For video datasets, the bind-time proposal MUST satisfy the DefaultFieldStrategy server constraint table (see the Video Dataset Field Constraints section above): video_url MUST be in IndexFields and MUST NOT be in FilterFields / SuggestFields; content_id, content_type, parent_content_id, and sequence_index MUST each be in FilterFields and MUST NOT be in / . Verify this table before rendering the Stage B dialog and before calling ; never submit a proposal that violates it, regardless of what the LLM inference or a user override suggests.