tray-gotchas

name	tray-gotchas
description	Load when debugging a Tray workflow error or working with known-weird surfaces — `raw_http_request` input format, `has_dynamic_output` rules, `trigger-reply`'s empty inputSchema, callable workflow sync/async pairing, jsonpath output-shape differences across step types, scheduled-trigger `day_of_week` explicit-keys rule, Salesforce custom (`*__c`) field picklist values, and stale `version_id` recovery.
disable-model-invocation	false

Tray Workflow Gotchas

Consult this when debugging workflow errors or working with complex structures.

Error Quick Reference

Error	Likely Cause	Fix
"Exported structure validation failed"	Malformed step properties, loop missing _loop wrapper, "next" keys in steps_structure	Check property type-wrapping. Use get_workflow to inspect current state
"Connector version not found"	Wrong version	Use list_connectors to get exact version
"Connector unexpectedly failed"	Wrong has_dynamic_output	Check operation's hasDynamicOutput field
"Required key [body] not found"	raw_http_request missing body	Set body explicitly even for GET
Workflow hangs	Missing callable response in sync mode	Add callable-workflow-response step
trigger-reply "not configurable"	Empty inputSchema misleads researcher	Use known property structure — see Trigger Event Reply section below
Silent step failures	Property values not wrapped	Use {"type": "string", "value": "..."} format
Missing connector icon in UI	Step name includes operation (e.g., slack-send-message-1)	Use connector name only: slack-1, never include the operation name
Stale version_id error	Used an old version_id after a mutating operation	Use the version_id from the error response (or from the last successful operation)
create_auth_collection fails with "service parameter is required"	Passed connector_name/connector_version instead of service UUID	Call list_service_environments first to get the service UUID and service_environment_id
list_service_environments fails with "service_name parameter is required"	Passed service_id instead of service_name	Use service_name (string, e.g., "google-gmail") and service_version (integer)
update_workflow_steps with error_handling returns 400	Omitted auth_id on an authenticated step entry	Always include auth_id when updating steps that have authentication, even if just setting error_handling
String contains {{ or template syntax	Used Handlebars/Mustache/moment.js — Tray has NO template engine	Use a script step to compose dynamic strings, then reference via jsonpath
Step output returns null/empty at runtime	JsonPath doesn't match the referenced step's actual output schema	Build jsonpaths from the referenced step's output schema — different step types have different shapes. Note: call_connector nests data under output, but the workflow runtime may not — always verify the runtime shape
Empty alert/email body inside an error branch	Used bare $.errors.message (Tray reads errors as a step name, needs the failed step's name next) or referenced the failed step's own output (empty on failure)	Use $.errors.<failed-step-name>.<field> — fields mirror the step's output schema in snake_case (e.g. $.errors.http-client-1.message, $.errors.http-client-1.status_code). The failed step's name is the second segment
Workflow "passed" but downstream jsonpaths return undefined at runtime	data passed to trigger_workflow wasn't shaped like $.steps.trigger (e.g. payload at root instead of nested under body for a webhook)	Shape data as the full trigger envelope: webhook → {body:{...}, query:{...}, headers:{...}}, callable-trigger → {data:{...}}, form-trigger → {result:{...}}, agent-tool-trigger → {tool_input:{...}}. Returned fired_payload_shape is the literal $.steps.trigger — verify jsonpaths against it

Runtime Debug Flow

When a workflow has run and you need to know why it succeeded or failed:

Permission first. trigger_workflow fires against the live Tray workspace — it executes every step, creates an execution log, consumes API quota, and runs every connector against whatever real system it's wired to. Always ask the user explicitly before firing, naming the side effects: "Ready to fire the workflow in your Tray workspace? This will create execution logs and consume API resources, and the Slack step will post to #demo-channel." Don't use vague language like "ready to test?" — be specific about what will happen.

1. Fire with trigger_workflow(workflow_id, data). Shape data exactly like $.steps.trigger should look at runtime — i.e. the trigger's full envelope. Examples:

   • webhook → data: {body: {<your fields>}, query: {}, headers: {}} (method/path/body_url/etc. are optional; Tray fills them on a real fire)
   • callable-trigger → data: {data: {<your fields>}}
   • form-trigger → data: {result: {<your fields>}}
   • agent-tool-trigger → data: {tool_input: {<your fields>}} or {static_data: {<your fields>}}

   The tool validates data against the trigger operation's outputSchema (the runtime envelope mirrored by $.steps.trigger — not inputSchema, which is for trigger configuration like add_raw_body/allow_cors/timeout) and rejects missing required keys with the declared envelope. Unknown top-level keys surface as warnings — most commonly a sign that you put the payload at root instead of nesting it (e.g. data: {timezone: 'X'} for a webhook should have been data: {body: {timezone: 'X'}}). The returned fired_payload_shape is the literal $.steps.trigger — verify your jsonpaths against it.

2. Call get_workflow_execution(workflow_id, execution_id) (or latest: true). Read failed_step_names and scan step_summary. Always check child_executions — a parent can report success while a callable failed silently.

3. For each failed step, call get_workflow_step_detail(workflow_id, execution_id, step_name). Step bodies are truncated to ~8 KB by default — bump max_body_bytes if the error message itself got cut.

4. Fix and re-fire. The same trigger_workflow returns a new workflow_instance.

If you need to find a recent execution to debug without firing, list_workflow_executions(workflow_id) returns a lightweight summary of the most recent runs.

The retry tiebreak in step-name resolution is verified against synthetic fixtures only — if a real retried step resolves to the wrong attempt, pass the explicit step_id from step_summary instead.

Critical Structural Rules

Loop Content Must Use _loop Wrapper

// WRONG - causes validation failure
{"name": "loop-1", "type": "loop", "content": [steps...]}

// CORRECT
{"name": "loop-1", "type": "loop", "content": {"_loop": [steps...]}}

Authentication on Steps

Use the auth_id parameter on each step entry passed to add_workflow_steps or update_workflow_steps — this sets metadata.auth_uuid on the step. Get the auth ID from list_authentications.

get_workflow (any view) surfaces step auths as metadata.auth_uuid: "<uuid>" — a bare string, ready to pass back as auth_id on update_workflow_steps. The view: "auths" shape returns the same UUIDs as a flat list per step, useful when you want to copy an existing auth onto a new step. (The write path wraps it as {value: "<uuid>"} internally; the read shape is bare.)

Step Ordering Is Array Position, NOT next Pointers

// WRONG - causes validation failure
[
  {"name": "trigger", "type": "normal", "content": {}, "next": "salesforce-1"},
  {"name": "salesforce-1", "type": "normal", "content": {}, "next": "script-1"}
]

// CORRECT - array order determines execution sequence
[
  {"name": "trigger", "type": "normal", "content": {}},
  {"name": "salesforce-1", "type": "normal", "content": {}},
  {"name": "script-1", "type": "normal", "content": {}}
]

All properties must be type-wrapped

See /tray-workflows:tray-connectors for the canonical type table (string, integer, number, boolean, array, object, jsonpath) and the jsonpath examples.

Object array items must also be wrapped

When the inputSchema declares items.type: "object" (or items.oneOf: [...]), each item in the array must be wrapped as {"type":"object","value":{...}}. A bare object inside an array round-trips through Tray's API but is unreadable by the workflow editor — the UI shows "field is required" / "value is required" on every nested property and the step does not fire at runtime. See /tray-workflows:build-workflow#object-array-items-must-be-wrapped-too for the full pattern. The MCP server validator rejects this on every mutation.

Common offenders to watch for:

• salesforce.find_records.conditions — items are {field, operator, value} objects.
• salesforce.update_record.fields — items are {key, value} pairs (note key, not name, despite the UI label).
• agent-tool-trigger.tool_input — input parameter array items.
• google-sheets/airtable column-mapping array fields.

Trigger public_url requirement

See /tray-workflows:tray-connectors for the internal-trigger table and which triggers need public_url.

has_dynamic_output Rules

• Check each operation's hasDynamicOutput from list_connector_operations
• DDL/private operations: usually false
• Data query operations: usually true
• Exception: raw_http_request always use false (fails with true)

What has_dynamic_output Actually Does

When using call_connector for testing:

• has_dynamic_output: true returns {output_schema: {...}} — the shape of the response for Tray's UI schema inference, NOT the actual data
• has_dynamic_output: false returns {output: {...}} — the actual data from the connector

For field discovery, you often want the actual data, so use false. For example, when discovering Salesforce fields via so_object_describe, use has_dynamic_output: false to see the actual field list — or use a SOQL query instead.

raw_http_request Input Format

• url: Must be object {"endpoint": {"type": "string", "value": "/path"}} or {"full_url": ...}
• parse_response: Must be string "true", not boolean
• body: Required even for GET — use {"none": {"type": "null", "value": null}}
• method: String — GET, POST, PATCH, PUT, DELETE

String Interpolation: brace-jsonpath only

Tray supports exactly ONE form of string interpolation: single-brace jsonpaths inside {type:"string", value:"..."} values. The runtime substitutes {$.steps.<name>.<path>} (and {$.env.x}, {$.config.x}) at execution time.

// CORRECT — single-brace, with `$`
"text": {"type": "string", "value": "Found {$.steps.salesforce-1.totalSize} opportunities for {$.steps.trigger.body.account_name}."}

// CORRECT — single-value reference (preserves source type)
"text": {"type": "jsonpath", "value": "$.steps.script-1.result"}

// WRONG — handlebars, expressions, raw $.steps without braces
"value": "Found {{salesforce-1.total}} records"
"value": "{{moment(trigger.datetime).toISOString()}}"
"value": "Found $.steps.salesforce-1.total records"

When to use which:

• One value, no static text → {type:"jsonpath", value:"$.steps.x.y"} (preserves type)
• Static prefix/suffix + one or more values → {type:"string", value:"...{$.steps.x.y}..."} (interpolation)
• Transformation, iteration, or computation → script step

Reach for a script step ONLY when transformation/iteration/computation is needed. Don't add a script for plain concatenation — brace interpolation handles that. See /tray-workflows:tray-connectors for the full pattern.

JsonPath Gotchas

• Build jsonpaths from the referenced step's actual output schema: Different step types have different output shapes. Always check the referenced step's output structure rather than assuming a universal pattern. Note: call_connector nests data under output, but the workflow runtime may structure it differently.
• Loop values: $.steps.loop-1.value.fieldName (not $.steps.loop-1.fieldName)
• Call workflow responses: $.steps.call-workflow-1.response.fieldName (not $.steps.call-workflow-1.fieldName)

Callable Workflow Rules

• trigger_and_respond operation MUST have a callable-workflow-response step
• trigger operation must NOT have a response step
• Synchronous callable workflows that don't return a response will hang

Trigger output shape — read it from outputSchema.properties

Trigger output paths are deterministic: the segment immediately after $.steps.trigger. MUST be one of the keys declared in the trigger operation's outputSchema.properties. No invention, no copying from a different trigger type. Confirmed envelopes:

Trigger	First segment after $.steps.trigger.
webhook	body (plus query, headers siblings)
form-trigger	result
callable-trigger	data (input fields land under here regardless of sync vs async invocation)
agent-tool-trigger	tool_input or static_data

The build-time validator (validate_workflow MCP tool) enforces this — paths like $.steps.trigger.body.<x> against a callable-trigger reject at build time with "available keys: data". If you see that error, the agent has used the wrong envelope key for the trigger type. The fix is always the same: read the trigger operation's outputSchema.properties and use one of those keys.

For unfamiliar trigger types (rare), call list_connector_operations(connector_name: "<trigger>", operation_name: "<op>") and look at the outputSchema.properties returned — never guess.

Trigger Event Reply Steps (trigger-reply)

The trigger-reply connector has an empty inputSchema, so its property shape is declared by the trigger operation's replySchema, not by trigger-reply itself. Read it with list_connector_operations(connector_name: "<trigger>", operation_name: "<op>") and shape properties to match — never memorise.

The shape rules split by trigger family (verified 2026-05-21 via direct schema capture + round-trip persistence, testing/local/RESEARCH-2026-05-21-replyschemas.md):

Trigger op	Reply shape
webhook.webhook_with_response	flat {status?, body?, headers?}
api-operation-trigger.request_response	{response: {<discriminator>: {...}}} — discriminated oneOf of 4 variants, dynamic inner
api-operation-trigger.publish_subscribe	flat {id?, eventType?, data} per event
agent-tool-trigger.request_response	{response: {<discriminator>: {...}}} — discriminated oneOf of 2 variants, static inner
replySchema: null (e.g. webhook.fire_and_forget, async callable-trigger.trigger, etc.)	no reply step allowed — reply_step_unsupported_trigger

Tray's workflow API does NOT enforce reply shape at write-time — the validator (1.12.0+) is the only static safety net. Garbage replies persist silently and only surface at runtime as ambiguous errors. Build-time correctness via the validator is therefore load-bearing.

Webhook — webhook.webhook_with_response (flat)

replySchema declares {status, body, headers} with required: [] and additionalProperties: false. Persist flat under properties:

"properties": {
  "status":  {"type": "number", "value": 200},
  "body":    {"type": "object", "value": {"ok": {"type": "boolean", "value": true}}},
  "headers": {"type": "array", "value": [
    {"type": "object", "value": {
      "key":   {"type": "string", "value": "Content-Type"},
      "value": {"type": "string", "value": "application/json"}
    }}
  ]}
}

required: [] means nothing is mandatory — return whichever subset of {status, body, headers} makes sense. additionalProperties: false means response, successBody, errorBody, or any other extra top-level key is rejected. Webhook is the trigger family where the no-envelope rule applies; the validator emits reply_shape_mismatch if you add one.

API operation trigger — request_response (discriminated oneOf, dynamic inner)

replySchema.properties.response is a oneOf of four variants, each requiring a single discriminator key:

Discriminator (in the reply)	Variant	Inner schema source
successBody	Success	trigger.properties.response.success.value
internalErrorBody	Internal Error	trigger.properties.response.internalError.value
invalidInputBody	Invalid Input	trigger.properties.response.invalidInput.value
forbiddenBody	Forbidden	trigger.properties.response.forbidden.value

Persist as {response: {<discriminator>: {...}}}. Exactly one discriminator per reply (oneOf):

"properties": {
  "response": {"type": "object", "value": {
    "successBody": {"type": "object", "value": {
      "id":   {"type": "number", "value": 42},
      "name": {"type": "string", "value": "widget"}
    }}
  }}
}

The inner shape of each <discriminator> is dynamic — defined by the agent when configuring the APIM trigger. For example, if trigger.properties.response.success.value = {type:"object", properties:{id:{type:"number"}, name:{type:"string"}}, required:["id","name"], additionalProperties:false}, then the reply's response.value.successBody.value must contain id and name, both wrapped with their declared types, and nothing else.

Critical: the DynamicSchema's .value is RAW JSON Schema, NOT type-wrapped. This is a single exception to Tray's "wrap every property value in {type, value}" rule. The properties tagged DynamicSchema in the trigger's inputSchema (APIM's request.body, request.headers, response.<variant>; agent-tool-trigger's body.tool_input, body.static_data) take an outer {type:"object", value:<schema>} wrapper as normal, but the <schema> IS a raw JSON Schema fragment — not recursively wrapped. The agent must write:

"body": {
  "type": "object",
  "value": {
    "type": "object",
    "properties": {
      "timezone": {"type": "string", "title": "Timezone"}
    },
    "required": ["timezone"],
    "additionalProperties": false
  }
}

and NOT:

"body": {
  "type": "object",
  "value": {
    "type":       {"type": "string",  "value": "object"},               // ❌
    "properties": {"type": "object",  "value": {"timezone": {...}}},    // ❌
    "required":   {"type": "array",   "value": [{"type":"string","value":"timezone"}]},  // ❌
    "additionalProperties": {"type": "boolean", "value": false}         // ❌
  }
}

Tray's workflow API accepts the wrong form silently (it stores anything), but the Tray UI rejects it as "Unexpected json schema format" and the schema is unusable for both UI rendering and DynamicSchema-based runtime validation. The reply-shape validator (1.12.0+) catches this as reply_shape_mismatch pointing at the trigger's broken DynamicSchema path.

Inside the raw JSON Schema fragment, each property's title populates Tray's UI "Name" field and description populates "Description" — include both whenever there's user-facing copy, otherwise the schema renders with a blank Name next to the key.

Build-time rules the validator enforces:

• response is required (envelope check).
• Exactly one discriminator key under response.value. Two or more → reply_shape_mismatch (multiple discriminators). Zero → reply_shape_mismatch (missing discriminator) with the allowed-keys hint.
• Unknown keys under response.value (not one of the four discriminators) → rejected.
• Inner walk: each discriminator's .value is validated against the trigger's matching response.<variantInputKey>.value JSON Schema (variantInputKey = discriminator without the Body suffix). Unknown inner keys, missing required inner keys, and leaf type mismatches all reject.

Don't leave trigger.properties.response.success.value empty when you build the trigger — the dynamic inner walk silently accepts everything if the schema is empty, leaving the contract ambient. Even a minimal {type:"object", properties:{}, additionalProperties:false} is better than nothing.

API operation trigger — publish_subscribe (flat per-event)

APIM's "Receive Events" operation — SSE-style streaming where each trigger-reply emits one event back to the caller. replySchema declares {id?, eventType?, data} per event with required: ["data"] and additionalProperties: false:

"properties": {
  "id":        {"type": "string", "value": "evt-1"},
  "eventType": {"type": "string", "value": "progress"},
  "data":      {"type": "string", "value": "{\"percent\":50}"}
}

data is required and typed string — serialize per-event payloads to strings (or use a jsonpath wrapper pointing at a script that produces a string). No response envelope, no discriminator — flat, like webhook.

Agent tool trigger — request_response (discriminated oneOf, STATIC inner)

replySchema.properties.response is a oneOf of two variants — materially different from APIM:

Discriminator	Variant	Inner schema
successBody	Success	type: ["object", "string"], additionalProperties: true — open: any object OR any string
errorBody	Error	{type: "object", properties: {message: {type: "string"}}, required: ["message"], additionalProperties: false}

successBody is open — the agent contract is whatever the tool returns:

"properties": {
  "response": {"type": "object", "value": {
    "successBody": {"type": "object", "value": {
      "customer": {"type": "object", "value": {
        "id":    {"type": "string", "value": "u-1"},
        "name":  {"type": "string", "value": "Ada"},
        "email": {"type": "string", "value": "[e-mail]"}
      }}
    }}
  }}
}

Or as a string:

"properties": {
  "response": {"type": "object", "value": {
    "successBody": {"type": "string", "value": "lookup complete"}
  }}
}

errorBody is fixed:

"properties": {
  "response": {"type": "object", "value": {
    "errorBody": {"type": "object", "value": {
      "message": {"type": "string", "value": "customer not found"}
    }}
  }}
}

Build-time rules the validator enforces:

• response required.
• Exactly one of successBody / errorBody under response.value. APIM's internalErrorBody / invalidInputBody / forbiddenBody do NOT exist here — they're rejected as unknown discriminators.
• errorBody must be {message: string} exactly — missing message, extra keys, or wrong leaf type all reject.
• successBody accepts any object (with any keys) or any string. No inner walk.

Key points

• Always read replySchema from list_connector_operations for the trigger op. Different trigger families, different shapes.
• "Flat with no envelope" is only true for webhook + publish_subscribe. APIM and agent-tool-trigger require the response: {<discriminator>: {...}} envelope.
• APIM has 4 discriminators with dynamic inner. agent-tool-trigger has 2 discriminators with static inner. Don't extrapolate one to the other.
• Webhook triggers split three ways: fire_and_forget returns 200 immediately (no reply step); webhook (validate-and-respond) auto-200s (no reply step); only webhook_with_response pairs with trigger-reply.

Callable Workflow Response Step (callable-workflow-response)

Pairs with callable-trigger.trigger_and_respond (the sync flavour) to return a value to the caller workflow. Unlike trigger-reply, the response shape is declared by the response connector's inputSchema, not by the trigger's replySchema (the callable trigger's replySchema is null).

callable-workflow-response.response.inputSchema declares one top-level key:

{
  "type": "object",
  "properties": {
    "response": {
      "type": "object",
      "description": "The response data to send to the caller workflow",
      "additionalProperties": false
    }
  },
  "additionalProperties": false
}

So the persisted property shape is, again, flat — a single response key wrapped as an object:

"properties": {
  "response": {"type": "object", "value": {"ok": {"type": "boolean", "value": true}}}
}

The per-step validator covers this on every add_workflow_steps / update_workflow_steps call — the cross-step validate_workflow audit does not need extra logic here.

Disabling Callable Workflows

Callable workflows can't be turned off directly. Tray prevents update_workflow_metadata({enabled: false}) on a callable to avoid breaking other workflows that depend on it. To stop a callable from doing real work, add a terminate step at the very top of its structure (before the trigger's actual children) — every invocation then short-circuits without running anything else.

Modifying Existing Workflows

Use get_workflow to fetch the current live state, then use update_workflow_steps to change step properties, remove_workflow_step / add_workflow_steps to add/remove steps, or update_workflow_structure to reorder steps or move them into/out of branch paths.

update_workflow_steps replaces the entire properties object for each targeted step — it is not a merge/patch. Always provide the complete properties per entry.

Scheduled Trigger

For the full daily-operation property shape (including the 7-key day_of_week object), see /tray-workflows:tray-patterns. The MCP validator (validate_workflow / per-step pre-flight) rejects scheduled-daily triggers missing any of hour / minute / tz / day_of_week, or day_of_week missing any of the 7 day keys.

Trigger output is zero-indexed for month and day_of_week — January is 0, Monday is 0. Conditions that compare these against a 1-indexed literal will silently route wrong at runtime. Use the Date Time Helper's Get Date Property operation rather than parsing the trigger payload directly.

Field Value Discovery

• UI display names often differ from API values (e.g., "Active" vs "4 - Current Customer (Active)")
• Custom fields have __c suffix
• Always use DDL operations or ask the user — never guess

Salesforce custom fields (*__c)

Whenever the user references a Salesforce field whose name ends in __c, the field is custom to that org and its valid values are not knowable from the connector's static schema. This applies to picklist values, lookup IDs, and free-text constraints alike.

Rule: before setting any value on a *__c field, you MUST do one of these:

1. DDL the picklist. If the field is a picklist, call the Salesforce DDL operation that lists its valid values (e.g., lookup_picklist_values_ddl or the equivalent for the version) and use a returned value verbatim.
2. Ask the user. "Just to confirm: is <value> a valid value for Priority__c in your org? It's a custom field so I can't verify on my own."
3. Surface the assumption explicitly. If the user has already supplied the value (e.g. "set Priority__c to High"), still call it out in the build summary: "Built with Priority__c = 'High' — this assumes 'High' is one of your existing picklist values for that field. If it isn't, the update step will fail at runtime."

Building silently with an unverified *__c value — even one the user gave you — is wrong. The cost of a one-line confirmation is tiny; the cost of a workflow that fails on every record at runtime is large.