webhook-template-builder（Webhook 模板生成器）

Webhook Template Builder

Purpose

Author and iterate Lytics webhook templates -- the server-side transforms that reshape a user profile into the body shape required by an arbitrary external webhook destination. The headline mode is build: the user names a destination (e.g. "Qualtrics") or pastes a docs URL, and the skill fetches the destination's API docs, infers the required payload + headers + auth model, drafts a template, iterates against /v2/template/{id}/test, saves, and emits a ready-to-go webhook job blueprint.

The skill never creates auths, connections, or jobs itself. After the template is saved, it hands off to connection-manager skill (auth) and job-manager skill (job) -- same orchestrator boundary as integration-setup skill.

Two webhook workflows are supported:

• webhook_triggers (default): event-based; fires on segment enter/exit.
• webhook_enrichment (--workflow=enrichment): request/response; sends a profile and lands the response into a stream.

Environment

Requires authenticated API access. See ../references/auth.md for credential resolution.

Invocation

list                                              # GET /v2/template
get <id-or-name>                                  # GET /v2/template/{id}
create <name> --type=js1|jsonnet|handlebars [--desired-format=<json-sample>]   # POST /v2/template (raw body); desired_format required for /test
update <id-or-name>                               # PUT /v2/template/{id}  (PUT, not POST -- API reference is wrong)
delete <id-or-name>                               # DELETE /v2/template/{id}
test <id-or-name> [--profile=<field>=<value> | --synthetic | --json=<file-or-paste>]
build <destination-name | docs-URL> [--workflow=trigger|enrichment]
draft <destination-name | docs-URL>               # alias for build that stops before save

Reads (list, get) execute immediately. Writes (create, update, delete) and the final save in build use the confirmation-gate pattern (../references/confirmation-gate.md).

Default Template Language

Default to js1 (JavaScript). The Lytics js1 runtime expects the entry point to be function template(data) {...} (NOT transformData, which appears in some external docs but does not match the runtime). Globals: ly_auth_config and ly_job_config (both JSON strings; JSON.parse them defensively -- see Step 5). Helper: hmacSha256(message, key) for signed requests.

Do not pivot languages without confirming

If js1 returns a non-deterministic error (e.g. /test 500 with empty body), do not silently switch to jsonnet. Diagnose first:

• Confirm desired_format is set on the template (see Probing Notes -- this is the most common cause of 500-with-empty-body on /test).
• Confirm the function is named template, not transformData.
• Test /test against an existing js1 template on the same account to rule out platform issues.

Only switch to jsonnet after surfacing the failure to the user and getting explicit approval. The switch rewrites the entire template body and the user must agree to the language change.

Use jsonnet when:

• The user explicitly requests it
• The destination needs a templating feature easier in jsonnet (heavy use of event.inSeg, event.segSlug, etc.)
• The user approved the switch after a confirmed js1 runtime issue

Research-Driven Authoring Flow (build)

Step 1: Identify Destination

Parse the user's input. Three forms:

• Name ("Qualtrics", "Slack")
• Docs URL ("https://api.qualtrics.com/...")
• Free-form description ("post a Slack message to #alerts when a user enters segment X")

If the input is a bare name, run a WebSearch for <destination> webhook events API reference and confirm the top result with the user before continuing.

Step 2: Fetch Destination Docs

Use WebFetch against the destination's API reference. Extract:

• HTTP method and URL pattern (note any path parameters)
• Required headers (auth header style, Content-Type, signature/HMAC headers if any)
• Request body shape -- a full JSON example, nested structure, required vs optional fields
• Auth model (API key in header, bearer token, HMAC-signed body, OAuth)
• Rate limits and non-retry status codes (so we can populate do_not_retry_for_http_status)

If the docs page is paywalled or empty, ask the user to paste the relevant section.

Step 3: Pull Sample Profile Context

Ground field references in real schema. Either:

• Real profile via entity-lookup skill (user provides identity field + value)
• Synthetic via schema-discovery skill (build a fixture from GET /v2/schema/user/field)

Use this to confirm Lytics field names referenced in the draft actually exist.

Step 4: Show Inferred Payload + Headers

Surface a summary the user can correct before we draft code:

Destination: Qualtrics XM Events API
Method: POST
URL pattern: https://{datacenter}.qualtrics.com/eventsubscriptions/{subscription}/events
Headers:
  - X-API-TOKEN: <secret>            -> needs auth provider
  - Content-Type: application/json
Body shape:
  { "events": [ { "type": "...", "user": { "email": "..." }, "data": { ... } } ] }
Auth model: API key in X-API-TOKEN header
Non-retry status codes: 400, 401, 403, 404

Surface anything the user must supply (datacenter, subscription IDs, secret values).

Step 5: Draft the Template

Generate js1 source using the runtime's expected entry point function template(data) {...}:

// Drafted by webhook-template-builder for <destination> on <YYYY-MM-DD>
function template(data) {
  // Defensive: ly_auth_config / ly_job_config may be undefined globals at runtime
  // (e.g., when /test is called with an empty job_config). Always guard with try/catch.
  var auth = {};
  var job  = {};
  try { if (typeof ly_auth_config === "string" && ly_auth_config) { auth = JSON.parse(ly_auth_config); } } catch (e) {}
  try { if (typeof ly_job_config  === "string" && ly_job_config)  { job  = JSON.parse(ly_job_config);  } } catch (e) {}

  return {
    events: [
      {
        type: "user_segment_enter",
        user: { email: data.email || "" },
        data: {
          first_name: data.first_name || "",
          last_name:  data.last_name  || "",
          ltv:        data.ltv        || 0
        }
      }
    ]
  };
}

Conventions:

• Function name is template, not transformData. The Lytics js1 runtime returns 500 with an empty body if the entry point is missing. Existing js1 templates on the platform consistently use template(data).
• Use data.<field> references for fields confirmed to exist in Step 3.
• Defaults: || "" for strings, || 0 for numbers, || [] for arrays. Webhook destinations rarely tolerate null.
• Always guard ly_auth_config and ly_job_config with typeof === "string" checks plus try/catch. The /test endpoint does not pass these globals when the corresponding key is empty/missing in the request, and an unguarded JSON.parse(ly_job_config || "{}") will throw ReferenceError at runtime.
• hmacSha256(message, key) if the destination signs requests.
• Header comment line for traceability.

Show the source to the user. Iterate on feedback in-memory; do not write to the API yet.

Step 6: /test Loop

The /test endpoint requires an existing template ID, so save a draft first. Use a draft name of the form _draft_<unix-ts>_<destination-slug> so it's easy to spot in list. desired_format must be set on create -- without it /test returns 500 with an empty body. Use the inferred response shape from Step 4 (a sample of what the destination expects); a minimal {} works as a fallback.

DRAFT_NAME="_draft_$(date +%s)_qualtrics"
# desired_format is a JSON sample of the destination payload; URL-encode it.
DESIRED_FORMAT=$(printf '%s' '{"firstName":"","lastName":"","email":"","embeddedData":{}}' \
  | python3 -c "import sys,urllib.parse; print(urllib.parse.quote(sys.stdin.read()))")

# Save the draft
curl -s -X POST \
  "${LYTICS_API_URL:-https://api.lytics.io}/v2/template?name=${DRAFT_NAME}&type=js1&description=Draft%20for%20${DESTINATION}&desired_format=${DESIRED_FORMAT}" \
  -H "Authorization: ${LYTICS_API_TOKEN}" \
  -H "Content-Type: text/plain" \
  --data-binary @template.js

Capture the returned template_id. Then test (see Test Workflow below). On feedback, update via PUT /v2/template/{id} and re-test. Repeat until the rendered output looks right.

Mention to the user that _draft_* templates may show up in their Lytics UI list; they will be renamed in Step 7.

Step 7: Save (Rename)

Once tests pass, prompt for the final name + description, then run update to rename the draft. Show the full payload through the confirmation gate (../references/confirmation-gate.md) before writing.

Step 8: Emit Job Blueprint and Hand Off

Print the full webhook job config payload (see Webhook-Job Hand-Off Blueprint below) with:

• template_id populated
• Inferred values (headers from Step 4, do_not_retry_for_http_status from rate-limit docs)
• Placeholders for the rest (auth_ids: ["<TODO via connection-manager>"], audiences: ["<segment_id>"])

Then explicitly hand off:

"Template saved as <final_name> (<template_id>). To wire up the job: 1. Run connection-manager skill to create a header_param_auth provider for the destination. Populate webhook_request_headers with key: value lines (one per line) covering everything the destination needs (X-API-TOKEN, Content-Type, etc.). Capture the returned auth_id. 2. Run job-manager skill with the config below, substituting the auth_id into auth_ids[]. Lytics auto-injects the auth's headers into every outgoing request at job runtime -- the job config itself does NOT have a headers field."

Test Workflow

The test request body is {job_config, entity}. Synthesize job_config from a stripped-down version of the eventual webhook job config (audience_trigger_events, webhook_url, headers, etc.) so the template runs against the actual config it'll see in production. The render output is a string -- that string is the body that would be POSTed to the destination.

Synthetic Fixture Defaults

For each type returned by GET /v2/schema/user/field, fill plausible values:

API Endpoints

List Templates

curl -s "${LYTICS_API_URL:-https://api.lytics.io}/v2/template" \
  -H "Authorization: ${LYTICS_API_TOKEN}"

List response is metadata-only -- no source body. Body is not retrievable from any GET endpoint (see Probing Notes); only name, type, description, desired_format, and timestamps are available.

Get Template

curl -s "${LYTICS_API_URL:-https://api.lytics.io}/v2/template/${TEMPLATE_ID}" \
  -H "Authorization: ${LYTICS_API_TOKEN}"

Returns metadata only (id, name, type, description, target, field_filter, desired_format, created, updated, author_id). The source body is not retrievable via any GET path -- see Probing Notes. If the user wants source-level history, recommend they keep the template body in their own version control.

Create Template

curl -s -X POST \
  "${LYTICS_API_URL:-https://api.lytics.io}/v2/template?name=${NAME}&type=${TYPE}&description=${DESCRIPTION}&desired_format=${DESIRED_FORMAT}" \
  -H "Authorization: ${LYTICS_API_TOKEN}" \
  -H "Content-Type: text/plain" \
  --data-binary @template.js

• Metadata (name, type, description, optional account_id, desired_format) goes in the query string.
• Template source goes in the body as raw text.
• type enum: js1, jsonnet, handlebars. The Lytics API reference page omits js1, but it is supported in production. If the API rejects it (400 invalid type), see Probing Notes.
• desired_format is required for /test to succeed. It is a sample of the destination's expected payload (a JSON string -- URL-encode it). If empty, /test returns 500 with an empty body. Use the inferred body shape from the destination's docs (Step 4); a minimal {} works as a placeholder if nothing better is available.

Update Template

curl -s -X PUT \
  "${LYTICS_API_URL:-https://api.lytics.io}/v2/template/${TEMPLATE_ID}?name=${NAME}&type=${TYPE}&description=${DESCRIPTION}&desired_format=${DESIRED_FORMAT}" \
  -H "Authorization: ${LYTICS_API_TOKEN}" \
  -H "Content-Type: text/plain" \
  --data-binary @template.js

Update is PUT /v2/template/{id}, not POST. The Lytics API reference page documents this incorrectly as POST; the live API responds to POST on this route with 405 Method Not Allowed and Allow: GET, OPTIONS, DELETE, PUT, HEAD. Don't "correct" this back to POST in a refactor.

Delete Template

curl -s -X DELETE "${LYTICS_API_URL:-https://api.lytics.io}/v2/template/${TEMPLATE_ID}" \
  -H "Authorization: ${LYTICS_API_TOKEN}"

Test Template

curl -s -X POST "${LYTICS_API_URL:-https://api.lytics.io}/v2/template/${TEMPLATE_ID}/test" \
  -H "Authorization: ${LYTICS_API_TOKEN}" \
  -H "Content-Type: application/json" \
  -d '{
    "job_config": {
      "audience_trigger_events": "enter",
      "webhook_url": "https://example.com/webhook"
    },
    "entity": {
      "email": "test@example.com",
      "first_name": "Ada",
      "last_name": "Lovelace",
      "_segments": ["high_value_customers"]
    }
  }'

Response: {data: "<rendered string>", status: 200, request_id: "..."}. The rendered string is the body that would be POSTed to the destination.

Test-endpoint quirks worth knowing:

• If the request body is missing job_config entirely, or job_config is {}, the runtime does NOT pass ly_job_config as a global. An unguarded JSON.parse(ly_job_config || "{}") then throws ReferenceError: ly_job_config is not defined. Always guard with try {if (typeof ly_job_config === "string" && ly_job_config) {...}} catch (e) {} (Step 5 scaffold does this).
• Runtime errors come back via {data: "<error message>", status: 200} -- HTTP 200 even when the template threw. Inspect data for the prefix "evaluating snippet: RUNTIME ERROR" (jsonnet) or a JS exception trace (js1).
• If /test returns HTTP 500 with an empty body, the most common cause is desired_format being unset on the template. Set it via PUT and retry.

Probing Notes

Several details of the template API drifted between the public reference docs and the live runtime. The findings below are confirmed against api.lytics.io. Treat them as ground truth; do not reorder or "correct" them based on docs alone.

Webhook-Job Hand-Off Blueprint

Tag-key conventions: ASK USER, INFER FROM DOCS, FROM SKILL OUTPUT, RESOLVE (probe Lytics).

How auth and headers are threaded

Webhook job configs do NOT have a headers field. Headers (X-API-TOKEN, Authorization, etc.) and URL query parameters live on the auth provider, not on the job. At runtime Lytics extracts them from the auth_ids[] reference and merges them into the outgoing request automatically.

When you hand off to connection-manager skill, instruct the user to create an auth provider of type header_param_auth and populate:

• webhook_request_headers -- newline-separated key: value text. Example for Qualtrics: X-API-TOKEN: <secret> Content-Type: application/json
• webhook_request_params -- URL query string format (key=value&key2=value2), if the destination needs query-param auth.

For OAuth (oauth2_client_credentials auth type), Lytics auto-injects Authorization: Bearer <token>. To customize, put a literal {token} placeholder (single braces, NOT {{token}}) inside the auth's webhook_request_headers -- e.g. Authorization: Token token={token} -- and Lytics replaces it server-side at request time.

Templates can ALSO read ly_auth_config (the auth's full config blob, JSON-encoded) inside the template body -- this is for HMAC signing or other body-side credential use, separate from header injection.

Verify the workflow config schema before emitting

The verified slugs as of 2026-04-27 are webhook_triggers and webhook_enrichment, with the field shapes shown in the blueprints below. GET /v2/provider returns the webhook provider but does NOT expose its workflows' config_routes on this API surface, so there's no clean machine-readable schema endpoint to probe. Instead, sanity-check against an existing webhook job in the destination account before emitting the blueprint:

# List webhook jobs and inspect their config shape
curl -s "${LYTICS_API_URL}/v2/job?show_all=true" -H "Authorization: ${LYTICS_API_TOKEN}" \
  | jq '.data[] | select(.workflow | test("webhook"; "i")) | {id, name, workflow, config}'

If at least one webhook job exists, its config keys are ground truth. If none exist, fall through to the blueprints below -- they were verified against a live webhook_triggers job and the underlying workflow definitions in lytics/lio (data/workflow/webhook_v2.json, data/workflow/webhook_enrichment.json, and the WebhookTriggers/WebhookEnrichment Go structs in src/api/v2/models/job_configs.go). Field names have drifted from public docs in the past (e.g., the live API uses audience_events not audience_trigger_events, segment_ids not audiences, do_not_retry as array[int] not a CSV string, and has no headers field at all) -- when in doubt, trust an existing live job over docs.

Audience Triggers (--workflow=trigger, default)

Corresponds to the webhook_triggers workflow. Verified shape:

{
  "name":        "<ASK USER -- defaults to '<destination> trigger for <segment>'>",
  "description": "<INFER>",
  "workflow":    "webhook_triggers",
  "auth_ids":    ["<FROM connection-manager hand-off -- the auth provider holds the credential headers>"],
  "config": {
    "segment_ids":         ["<ASK USER -- segment hex id(s) (NOT slugs)>"],
    "webhook_url":         "<INFER FROM DOCS or ASK USER -- must be HTTPS>",
    "http_method":         "POST",                                  // enum: GET|POST|PUT|PATCH|DELETE
    "audience_events":     "",                                      // enum: "" (both, default), "enters_only", "exits_only"
    "template_id":         "<FROM SKILL OUTPUT -- omit for raw default Lytics payload>",
    "user_fields":         null,                                    // optional: subset of profile fields to export; null = all
    "field_changes":       [],                                      // optional: up to 75 field names whose changes also trigger
    "withbackfill":        false,                                   // include users already in audience
    "include_segs":        false,                                   // include profile's segment membership as `segments_all`
    "batch_requests":      false,
    "batch_size":          10,                                      // max 1000
    "batch_flush_duration": 5,                                      // seconds, max 300
    "worker_count":        5,                                       // max 20
    "do_not_retry":        [400, 401, 403, 404],                    // array[int] -- HTTP statuses to NOT retry
    "cloudevents":         false,                                   // wrap payload in CloudEvents v1.0 envelope
    "cloudevents_type":    null,                                    // default: com.lytics.audience.user.event
    "cloudevents_source":  null                                     // default: /work/id
  }
}

User Enrichment (--workflow=enrichment)

Corresponds to the webhook_enrichment workflow. Verified shape:

{
  "name":        "<ASK USER>",
  "description": "<INFER>",
  "workflow":    "webhook_enrichment",
  "auth_ids":    ["<FROM connection-manager hand-off>"],
  "config": {
    "segment_ids":     ["<ASK USER -- segment hex id(s)>"],
    "stream":          "webhook_enrichment",                        // optional; lands the response in stream of this name
    "webhook_url":     "<INFER FROM DOCS or ASK USER -- must be HTTPS>",
    "http_method":     "POST",                                      // enum: GET|POST|PUT|PATCH (no DELETE for enrichment)
    "template_id":     "<FROM SKILL OUTPUT>",
    "user_fields":     null,                                        // optional: subset of profile fields to send
    "field_changes":   [],                                          // optional: trigger on field changes (up to 75)
    "withbackfill":    true,                                        // NOTE: defaults to true for enrichment (false for triggers)
    "audience_events": "",                                          // enum: "" (both) | "enters_only" | "exits_only"
    "worker_count":    5,                                           // max 20
    "do_not_retry":    [400, 401, 403, 404]                         // array[int]
  }
}

Annotate any field the skill could not infer with "review carefully" so the user catches it before job-manager writes. Note: there is no headers field on either workflow's config -- credentials are threaded via auth_ids[]. See "How auth and headers are threaded" above.

Behavior

For Read Operations (list, get)

Execute immediately. Display templates in a table format showing: ID, name, type, description, updated.

For Write Operations (create, update, delete, save step of build)

Use the confirmation-gate pattern. Show the full payload (raw source plus query-param metadata) and the resolved curl invocation; require explicit yes before writing.

For test

No confirmation needed for a test against an already-saved template -- it's read-only on the destination. Render the response inline. If the test was run against a draft created during build, treat the draft itself as having gone through the create gate already.

Error Handling

• 404 on template -- offer a fuzzy-match against list (case-insensitive substring on name).
• 405 on POST /v2/template/{id} -- the update verb is PUT, not POST. Retry with PUT.
• 500 with empty body on /test -- almost always means desired_format is unset on the template, OR the js1 entry point is named something other than template. Check both via GET /v2/template/{id}; set desired_format via PUT if missing; rename the function in the source if it isn't template(data).
• HTTP 200 with data: "ReferenceError: ly_job_config is not defined" -- the template parses an unguarded JSON.parse(ly_job_config). Wrap in try {if (typeof ly_job_config === "string" && ly_job_config) {...}} catch (e) {} and re-test.
• HTTP 200 with data: "ReferenceError: <field>" (jsonnet) or a JS exception trace (js1) -- referenced entity field missing. Add a default fallback or pick the right field name from Step 3 schema discovery.
• 400 on create with type=js1 -- API may not accept that enum on this account. Surface a clear error to the user; do not silently switch to jsonnet (Default Template Language section).
• 415 on create / update -- Content-Type probe failed; retry with the alternate (application/javascript for js1, application/jsonnet for jsonnet). Cache the working value for the session.
• HMAC mismatch reported by destination -- secret in ly_auth_config differs from the destination's expectation, or hmacSha256 is being passed inputs in the wrong order/encoding. Surface the destination's error verbatim.
• WebFetch returns no useful content (paywall, rate limit, JS-rendered) -- prompt the user to paste the relevant docs section.
• Destination returns 401 in the live job -- this skill saved a working template; the issue is auth. Route the user to connection-manager skill.

Dependencies

• Composes: entity-lookup skill, schema-discovery skill, connection-manager skill, job-manager skill
• Uses: ../references/auth.md, ../references/api-client.md, ../references/confirmation-gate.md, ../references/api-response-format.md
• Related: integration-setup skill (orchestrator hand-off pattern), integration-advisor skill (research-driven advisor pattern)