Testdata Generation

Test Data Generation

This directory is the skill. It has no baked-in company knowledge. Platforms are adapters; local files work with zero extra infrastructure.

This skill never invents business data. It decides which path to take and extracts parameters the user already supplied; an executor, a published tool, or a generated script is what actually calls the backend. "Construct succeeded" means the backend returned a business ID, not that an ID appeared in the reply.

SKILL_DIR is the folder that contains this SKILL.md. Every command below uses it, so resolve it once:

SKILL_DIR="<directory of this SKILL.md>"

Config lookup: $DATA_BUILD_CONFIG → ./testdata/config.yaml → ~/.testdata/config.yaml.

Four fallback paths, in this order, after the case-material check below:

1. Domain skill — search for a matching data-build skill, install or load it, and delegate
2. Existing tool — reuse a published tool from the tool registry
3. API catalog — discover the APIs that can build the data
4. Generated script — write a construction script from those APIs and run it

Documents (load on demand)

Read this file first; it is the routing contract. Load anything else only when the row below applies, so a single-step construct does not drag the whole tree into context.

Case-material route (check first)

After receiving a data-construction request, first decide whether it is a case-material job.

Trigger phrases (any match enters the sub-skill and stops the generic flow):

• English: case data, case materials, test-data preparation, write back test data, data requirements, data inventory, prepare test data, test material list
• Chinese: 用例数据, 用例物料, 用例数据准备, 为用例准备数据, 用例前置数据, 测试数据回写, 用例数据构造, 测试数据分析, 测试物料, 数据需求分析, 测试数据准备, 数据物料, 需求测试数据, 准备测试数据, 测试物料清单, 数据诉求

On hit: load and follow references/case-data-material-planner/SKILL.md. Run scripts/pipeline.ts as the orchestrator — construction, binding, and writeback are stages inside that script, so do not spawn Agents to do them.

The sub-skill root is <this SKILL.md directory>/references/case-data-material-planner/. All of its scripts/, agents/, templates/ paths are relative to that root.

Decision tree

Handle every other request in this exact order:

user data-construction request
    │
    ▼
[route] case-material scene? (phrases above)
    │
    ├─ yes → load references/case-data-material-planner/SKILL.md  ✓ stop
    │
    └─ no → generic flow
            │
            ▼
        [step 1] search_data_build.ts
                 (semantic proven methods + skill marketplace, in parallel)
            │
            ├─ pinned_matches present AND Agent judges them relevant
            │       → load skillPath (or install) → delegate → report  ✓ stop
            │
            ├─ proven_matches present AND Agent judges them applicable
            │       → run proven_invocation → feedback only  ✓ stop
            │
            ├─ skill_matches: one local clear match
            │       → load → delegate → report  ✓ stop
            │     several / weak / needs install → ask, then same
            │
            └─ none match
                    │
                    ▼
                [step 2] tool_registry.query (two query angles)
                    │
                    ├─ one clear tool (or user already gave an id)
                    │       → query_input_list → fill params → execute → report  ✓ stop
                    │     several / weak → ask, then same
                    │
                    └─ no matching tool
                            │
                            ▼
                        [step 3] api_catalog (in parallel when useful)
                            │
                            ├─ method A: plan change APIs (if planId exists)
                            └─ method B: keyword / service search
                            │
                            ▼
                        [step 4] write script from template
                            │
                            ├─ generate → run locally → report
                            └─ after a successful run, ask whether to publish
                                    ├─ yes → tool_registry.publish  ✓ stop
                                    └─ no  → return the result     ✓ stop

Step 1 — Search matching data-build options

Build the search input

Extract two kinds of input from the user request:

Keywords (skill-name match):

• Take 1–2 nouns that name the domain slot, such as the name in slot.yaml
• Never use verb phrases such as "create a" or "help me construct"
• If keywords miss, the script falls back to a full scan. Do not retry by hand.

Structured query (semantic match):

• registry-key: {entity}::{action}, e.g. catalog-order::create
• query: one natural-language sentence
• entry-type: entity or action
• domain: from workspace context business_line, or infer from the request

Examples and the full command: references/workflow.md.

node "$SKILL_DIR/scripts/search_data_build.ts" \
  --keywords <nouns> \
  --query "<sentence>" \
  --registry-key "<entity>::<action>" \
  --entry-type entity \
  --domain "<domain>" \
  --json

The script runs proven-method search and skill-marketplace search in parallel, then injects pinned favorites. JSON shape:

{
  "pinned_matches": [...],
  "proven_matches": [...],
  "skill_matches": [...]
}

Result priority

Priority 0 — pinned_matches (highest)

Pinned skills are returned unconditionally so retrieval noise cannot drop them.

The Agent judges relevance from each item's description (same idea as proven applicability):

• Relevant → use it first. Prefer skillPath when SKILL.md exists. If available=false or skillPath is missing, install then load.
• Not relevant → ignore and continue to proven / skill matches.

The script does no keyword filter on pinned items. A relevant pinned hit ends step 1; do not evaluate proven/skill after that.

Priority 1 — proven_matches

Reuse only when all three hold:

1. Semantic alignment: registry_key describes the same operation
2. Tool reachable: the bound skill is installed or the bound tool can execute
3. Params available: required proven_invocation.paramMapping values exist in context

All three → execute proven_invocation and keep experience_id for feedback. Any miss → skip that row and continue to skill_matches.

Priority 2 — skill_matches

Asking is for ambiguity, not ceremony. The user already asked to construct data. Drop rows whose description does not cover the request (a catalog-product construct is not a distributor slot, even if both appear in skill_matches).

• One remaining local match (skillPath already has SKILL.md) → load it, say which skill you used, and stop. Do not ask first.
• Needs install (available=false or no skillPath) → ask, because install is a side effect the user did not request.
• Two or more remaining rows that could both be right, or only a weak/partial match → show them and ask: "Which of these skills matches your request, if any?"

None match → go to step 2.

If search is empty and the user wants a reusable domain pack (not a one-off construct), read $SKILL_DIR/slot-scaffolder/SKILL.md and follow it. That is a bundled folder in this skill, not a separately installed skill — do not go looking for it in the host's skill list. One-off work continues at step 2.

Install and load a confirmed skill

SKILLS_DIR="$(dirname "$SKILL_DIR")"
node "$SKILL_DIR/scripts/adapters/cli.ts" skill_marketplace.install '<name>' '$SKILLS_DIR'

If skillPath already points at a directory that contains SKILL.md, load it directly and skip install. After load, follow that skill and stop.

Pin favorites (personal cheat-sheet)

Pinned skills always enter the candidate set first. Relevance is Agent-judged.

Storage: project ./testdata/data-build-favorites.json, user ~/.testdata/favorites.json. add writes the user file unless --scope project. $DATA_BUILD_FAVORITES_PATH overrides both.

Step 2 — Tool registry (only when no skill matched)

Query with natural-language sentences, not space-separated keywords.

Write at least two angles:

node "$SKILL_DIR/scripts/adapters/cli.ts" tool_registry.query "create a catalog test product"

If the user already has a tool id, look it up exactly:

node "$SKILL_DIR/scripts/adapters/cli.ts" tool_registry.get "<resource-id>"

• User already named a tool id, or exactly one result clearly matches → use it. Tell the user which tool you picked.
• Several results or only a weak match → show them and ask: "Is there a tool that matches your request?"

Then:

1. tool_registry.query_input_list(resource_id)
2. Fill params. Order IDs, user IDs, and other core business fields must be confirmed with the user. Do not invent them. Optional fields may use sane defaults.
3. tool_registry.execute(resource_id, params)
4. On success → report (see below)

Step 3 — API catalog

Method A — plan change APIs (preferred when a plan exists)

Use when the request sits inside a test-plan / change scope.

Resolve planId from workspace context testPlan.id or planId. If missing, ask the user.

node "$SKILL_DIR/scripts/adapters/cli.ts" api_catalog.search_plan_changes "<planId>"

Then api_catalog.detail(<operationId>) for key operations.

Method B — keyword / service search (fallback)

Use when there is no planId, or plan APIs are weakly related.

First: read workspace context targetRepositories[].serviceId or relatedJobs[].serviceId. If a service id is already known, list its APIs directly:

node "$SKILL_DIR/scripts/adapters/cli.ts" api_catalog.list_by_service "<serviceId>" "<optional-name-filter>"

Second: keyword search. Use one precise token, not a long Chinese/English phrase.

node "$SKILL_DIR/scripts/adapters/cli.ts" api_catalog.search "create catalog product"

Then api_catalog.detail(<operationId>).

If the workspace has no API source at all (no OpenAPI files, no planId, no serviceId), stop and ask for a plan, a service id, or an OpenAPI directory. That is a material gate, not a field-by-field param review. Do not invent endpoints.

Combination

Step 4 — Write and run a construction script

Write ./testdata/<name>.ts from references/script-template.ts, then run it.

Required shape: shebang + docstring, top-level constants, main(params) -> {success, data, error}, CLI entry via import.meta.url. Use adapter helpers only: callHttp, callSql, getConfig, callFeatureFlag.

Rules:

• Child-process calls use argument lists, never shell: true
• Timeouts on every network call
• Coerce object / array / bool / number with typeof / Array.isArray
• Do not invent order IDs, user IDs, or amounts

Feature flags (only when the user asks for experiments, drafts, or whitelists and feature_flags.type is not noop): confirm the environment before writes; production writes need a second confirmation; one subject per call. toolType for a successful flag script is feature_flag.

After a successful local run, ask whether to publish to the tool registry. Do not publish before a successful run.

node "$SKILL_DIR/scripts/adapters/cli.ts" tool_registry.publish '<name>' '<what the script does>' './testdata/<name>.ts' '[{"name":"<param>","type":"string","required":true}]'

What to tell the user

Report the business outcome, because that is the only part the user can act on. Lead with the identifiers the backend returned, then how they were produced, then anything still needed:

Constructed <what>: <field>=<value>, <field>=<value>
Path: <domain skill | registry tool | generated script> (<name>)
Next: <what the user can do with it, or what is still missing>

Keep infrastructure out of it — skillRoot, absolute paths, mock ports, node … invoke lines, and adapter names are noise to the person who asked for data, and in a case document they are actively wrong (see the writeback rules). Also say which environment produced the IDs when it was the local mock, since a mock ID looks identical to a real one but nothing landed in a real system.

If a path failed and you fell through to the next one, say so in one line rather than narrating every attempt.

Success follow-up (after any successful path)

Non-blocking. A report/feedback failure must not hide a successful construct.

When to report vs feedback:

• Step 1 proven hit that ran successfully → feedback only, do not report again
• Step 1 skill delegate succeeded → report
• Step 2 tool execute succeeded → report
• Step 4 script succeeded → report
• Step 4 published after a successful run → report with the published resource id

Payloads and toolType mapping: references/workflow.md.

Never put tokens, cookies, personal identifiers, or local absolute skillRoot paths in the experience store. Consumers locate a skill by resourceId (skill name).

Pin ask (only after a step-1 skill success)

If a skill (not a tool or script) completed the request and is not already pinned, ask: "Pin <name> so later similar requests stay at the top?"

On confirm:

node "$SKILL_DIR/scripts/favorites.ts" add --uuid <uuid> --name "<name>" \
  --desc "<what it does and when to use it>" --path "<local skill dir>"

Pinning and experience reporting are independent. A declined pin does not undo the construct.

Material gates (high-level only)

Ask the user only when a prerequisite material is missing — a case pack, an API/OpenAPI source, a cross-domain product, or a slot spec. Do not stop to collect executor fields (fulfillOn, quantity, credits, rate, …). Those follow the original rules: reuse user-supplied IDs, fill optional params with defaults, do not invent core IDs.

After the user supplies the material, resume and complete the remaining workflow. Do not skip later high-level steps.

Guardrails

• Do not invent business identifiers.
• Do not call organization-specific CLIs or private registries from this skill.
• SQL is SELECT-only.
• Feature-flag writes require an explicit environment and user confirmation.
• Keep this file under 500 lines; load references/workflow.md on demand.