Endpoint specification and error codes for the System Integration API of the Marketplace edition.
note
This page is the reference for the APIs called with a system integration (service account) token. A system integration can use Chat, Conversation Management, File Management, Artifact Retrieval, Search, Registered Agents and nine project APIs (Projects). For the APIs called with an application (sign-in from an external tool) token, see the User API Reference (Conversations, Organizations) and the User API Reference (Projects). The Library, Library Agent, Skills, Playbooks, Persona, User Management, Logs and Shared Memory APIs are available in the SaaS edition only. See the User API Reference (SaaS) and the System Integration API Reference (SaaS) for those.
Every API request must include a Bearer token in the Authorization header.
Authorization: Bearer <access_token>
If the token is missing or invalid, 401 Unauthorized is returned. The Marketplace edition applies no scope restrictions.
Call the APIs on this page with a service account token. The APIs that can be called with an application (sign-in from an external tool) token are in the User API References. Calling them with a user token obtained in any other way (such as the token used when signing in to the Main Screen) returns 404 Not Found for every endpoint.
Cursor value for the next page. Pass the cursor from the previous response as is.
order
string
Sort order: "asc" or "desc" (default). Any other value returns 400 Bad Request. Supported on some endpoints only.
Responses include a hasMore field. While it is true, pass the cursor value in the next request to retrieve the following page.
Cursor-based pagination is available on the list endpoints of Conversations, Messages, Files, Message search and Projects. Treat cursor as opaque: do not interpret it, just pass the previous value back.
Array of message objects. At least one user-role message is required. Even if multiple user messages are included, only the last one is processed. Conversation history is managed on the server side, so you do not need to send past exchanges in the messages array. To continue an existing conversation, specify the previous conversation ID in the conversationId parameter.
stream
boolean
必須
Specify true. Currently, only streaming mode is supported.
conversationId
string
—
Specify a conversation ID to continue an existing conversation. If omitted, a new conversation is created.
Use "user". "assistant" / "system" are accepted but only "user" is effectively used as the starting point of the agent turn in the current implementation
content
string | array
必須
The message content. A string, or an array of content parts (text / input_file). The text to be processed must be 1 MB or less in UTF-8.
2. data: {"choices":[{"delta":{"content":"..."}}]} × N
3. data: {"choices":[{"finishReason":"stop"}]}
4. data: [DONE]
Responses are delivered via SSE (Server-Sent Events) streaming. Each chunk is in OpenAI-compatible ChatCompletionChunk format. For details on the connection lifecycle, response objects, agent extension fields, and error handling, see the Streaming Guide.
Errors before the SSE stream starts are returned in JSON format. Errors during the stream are delivered as SSE chunks.
Status
Description
400
Invalid request parameters (missing or malformed model / messages / stream / conversationId / role / content, or more than 20 input_file parts in a single request). A malformed JSON body (code: "invalid_body"). Or the specified conversationId is archived (code: "conversation_archived")
400
The messages array has no message with the user role (code: "no_user_message")
415
The body Content-Type is not application/json (code: "unsupported_media_type")
403
The specified conversation is a project conversation (code: "project_conversation". Messages cannot be sent to a project conversation)
404
The fileId of the specified input_file was not found (code: "file_not_found")
409
The specified conversation is currently being processed (code: "conversation_processing")
GET/v1/chat/completions/:conversationId
Reconnect to the SSE stream of a conversation in progress. Use this when the agent is processing a long-running task or to recover after a network disconnection.
import requests
response = requests.get(
"https://<your-domain>/api/v1/chat/completions/550e8400-e29b-41d4-a716-446655440000",
headers={"Authorization": f"Bearer {token}"},
stream=True,
)
for line in response.iter_lines():
if line:
print(line.decode("utf-8"))
Reconnects to the SSE stream. The response format is the same as POST /v1/chat/completions. The response headers include X-Conversation-Id and X-Message-Id. For detailed reconnection procedures, see the Streaming Guide.
Invalid conversationId, messageId is not a string, the body is malformed JSON or not a JSON object (code: "invalid_body"), or the conversation is archived (code: "conversation_archived")
400
The message of messageId does not belong to the specified conversation (code: "conversation_mismatch")
400
The endpoint of this conversation does not support cancellation (code: "unsupported_endpoint"), or cancellation is not enabled in this environment (code: "post_processing_not_enabled")
415
The body Content-Type is not application/json (code: "unsupported_media_type")
404
Conversation not found, or the message cannot be canceled (code: "message_not_found"). message_not_found is returned in two cases: the message was just sent and could not be canceled after waiting about 10 seconds (sending again in a few seconds cancels it), or the message of messageId does not exist or is not yours (sending again does not cancel it; check messageId)
404
The message of messageId is not being processed (it has already finished) (code: "not_processing")
Lists every registered agent you can run in chat (agentId of POST /v1/chat/completions), the active ones only. The same agents also run in MCP (agentId of execute_chat) where the MCP endpoint is enabled and you use the chat toolset (?toolsets=chat, the default), and in A2A (/v1/a2a/agents/{agentId}/) where the A2A endpoint is enabled. It is the same list as the MCP list_agents tool and can be up to 5 minutes old. In chat, pass marketplaceAgentId as agentId.
Note: The Marketplace edition does not require scopes (authentication only).
Archived conversations are deleted automatically once the retention period set by an administrator has passed since they were archived. Their messages, and the attached and artifact files used only by that conversation, are deleted with them. If no retention period is set, they are not deleted.
Conversations retrieved with mode=archive only support read operations (list / message history) and restoration. The following operations return a 400 error (code: "conversation_archived").
Operation
Behavior on archived conversations
GET /v1/conversations?mode=archive
✅ Allowed
GET /v1/conversations/:id/messages
✅ Allowed (messages[].archived: true is added)
PATCH /v1/conversations/:id
❌ 400 Archived conversations cannot be modified
DELETE /v1/conversations/:id
❌ 400 Archived conversations cannot be deleted via this endpoint
POST /v1/chat/completions (continuing an existing conversation)
❌ 400 Cannot send messages to an archived conversation
POST /v1/chat/completions/:id/cancel
❌ 400 Cannot cancel an archived conversation
POST /v1/conversations/:id/restore
✅ Restore to active
GET/v1/conversations
List conversations. Supports cursor-based pagination.
Cursor for the next page (the cursor value from the previous response)
order
string
—
"desc"
"asc" or "desc"
title
string
—
—
Filter by title
mode
string
—
"active"
"active" (regular conversations) or "archive" (archived conversations). Switching `mode` switches to a different data store, so a `cursor` cannot be carried over
Kind of conversation: "chat" (chat) or "mail" (started from mail). Conversations of project steps are not in the list (their messages can be read by specifying the conversation ID). Not returned with mode=archive
conversations[].createdAt
string
Creation date and time (ISO 8601)
conversations[].updatedAt
string
Last updated date and time (ISO 8601)
conversations[].archived
boolean?
Whether archived. Always true when retrieved via mode=archive. When mode=active returns this field as true, it is only because the conversation remained in archive state but was returned by the active list — usually omitted
hasMore
boolean
Whether the next page exists
cursor
string
Cursor for fetching the next page. Included only when hasMore is true
limit is invalid (outside 1-100), mode is not "active" / "archive", order is not "asc" / "desc", cursor is invalid (not a value from the previous response), or title is not a string (for example, it was specified more than once)
502
upstream_connection_error
The list of conversations could not be retrieved (retry after a while)
GET/v1/conversations/:conversationId/messages
Retrieve the message history of a specified conversation. Cursor-based pagination is supported. Retrieving it does not mark the conversation as read (the same goes for reading via MCP or A2A). To mark it as read, call Mark a conversation as read.
The message body text. If the body is empty, the result (content) of the latest task is returned
messages[].createdAt
string
Creation date and time (ISO 8601)
messages[].updatedAt
string
Last updated date and time (ISO 8601)
messages[].status
object
Message processing status
messages[].status.processing
boolean
Whether the agent is currently processing
messages[].status.unfinished
boolean
Whether the agent response was interrupted
messages[].tasks
array
Array of tasks executed by the agent (see below)
messages[].deliverables
array
Array of artifacts generated by the agent. Included only when present
messages[].deliverables[].filename
string
Artifact file name
messages[].deliverables[].filepath
string
Artifact file path
messages[].deliverables[].source
string
"agent" (agent-generated) or "user" (user-uploaded)
messages[].deliverables[].isPrimary
boolean
Whether it is a primary artifact
messages[].deliverables[].createdAt
string
Creation date and time (ISO 8601 format). An empty string when unknown
messages[].deliverables[].mimeType
string
MIME type (e.g., "application/pdf", "text/markdown"). "application/octet-stream" when it cannot be determined
messages[].deliverables[].size
number?
File size (bytes). Omitted when 0 or below
messages[].files
array?
Array of user-uploaded files. Included only when present for messages with sender: "User"
messages[].files[].fileId
string
Unique identifier of the file (UUID)
messages[].files[].mimeType
string
MIME type (e.g., "image/png", "application/pdf")
messages[].files[].width
number?
Image width in pixels. Only for image files
messages[].files[].height
number?
Image height in pixels. Only for image files
messages[].archived
boolean?
true only for messages retrieved via archive. Added when viewing the message history of a conversation retrieved with GET /v1/conversations?mode=archive
hasMore
boolean
Whether a next page exists
cursor
string
Cursor for the next page. Included only when hasMore is true
Task execution result text. Omitted when not present
metadata
object?
Auxiliary information that varies by actionType (sub_event_type, url, filepath, etc.). See the Streaming Guide for details
files
array?
Files related to the task (filename / size / filepath / mimeType?. mimeType is omitted when it cannot be determined). The array itself is omitted when there are no related files
timestamp
number
Timestamp of task execution (Unix milliseconds)
createdAt
string
Creation date and time (ISO 8601)
updatedAt
string
Last updated date and time (ISO 8601)
Safely ignore unknown fields
Future implementations may add new fields to actionType / metadata. Client implementations should safely ignore any fields not listed in this table. For detailed metadata structure, refer to the Streaming Guide — Task data model.
# First page
curl "https://<your-domain>/api/v1/conversations/<conversation_id>/messages?limit=50" \
-H "Authorization: Bearer <access_token>"
# Second page (use the cursor returned by the first page)
curl "https://<your-domain>/api/v1/conversations/<conversation_id>/messages?limit=50&cursor=<cursor>" \
-H "Authorization: Bearer <access_token>"
If limit is omitted, 50 items are returned. To fetch everything, keep following the cursor until hasMore becomes false.
Using the cursor
The cursor represents a position in the result set. Do not change order while you are following the same cursor — changing the conditions changes the target set and causes duplicated or skipped items.
Items may also be duplicated or skipped across pages if messages are added or removed while you are paging.
conversationId is the all-zero UUID (00000000-0000-0000-0000-000000000000). This value is not accepted
400
missing_parameter
title is missing
400
invalid_title
title is an empty string, is not a string, or exceeds 200 characters
400
invalid_body
The body is malformed JSON
400
conversation_archived
The conversation is archived
415
unsupported_media_type
The Content-Type of the body is not application/json
404
conversation_not_found
The conversation was not found or is not yours
DELETE/v1/conversations/:conversationId
Delete a conversation. Of the files attached to its messages and its artifact files, those not used by any other conversation are deleted too (they no longer appear in GET /v1/files).
conversationId is the all-zero UUID (00000000-0000-0000-0000-000000000000). This value is not accepted
400
conversation_archived
The conversation is archived. Use POST /v1/conversations/:id/restore to restore it
404
conversation_not_found
The conversation was not found or is not yours
409
conversation_processing
The conversation cannot be deleted because it is currently being processed
409
project_conversation
A project conversation. Project conversations cannot be deleted
409
legal_hold
A conversation under legal hold. It cannot be deleted until an administrator releases the hold
POST/v1/conversations/:conversationId/restore
Restore an archived conversation back to active. After restoration, the conversation can be edited, continued, and deleted just like a regular conversation.
conversationId is the all-zero UUID (00000000-0000-0000-0000-000000000000). This value is not accepted
404
conversation_not_found
The conversation does not exist in the archive, or is already active (not distinguished from the archived state)
POST/v1/conversations/:conversationId/read
Marks a conversation as read. Reading the messages of a conversation through the API (REST, MCP, or A2A) does not mark it as read, so call this when you want to record that it was read. The read position never moves back (specifying an earlier message does not move it before where it is).
The body can be omitted. When you send one, use Content-Type: application/json.
パラメータ
型
必須
説明
messageId
string
—
ID of the last message you read. If omitted, the conversation is marked as read up to its latest message at that moment (specify it to avoid marking messages that arrived while you were reading)
# First page
curl "https://<your-domain>/api/v1/files?limit=50&order=desc" \
-H "Authorization: Bearer <access_token>"
# Second page (use the cursor returned by the first page)
curl "https://<your-domain>/api/v1/files?limit=50&order=desc&cursor=<cursor>" \
-H "Authorization: Bearer <access_token>"
If limit is omitted, 50 items are returned. To fetch everything, keep following the cursor until hasMore becomes false.
Using the cursor
The cursor represents a position in the result set. Do not change order while you are following the same cursor — changing the conditions changes the target set and causes duplicated or skipped items.
Items may also be duplicated or skipped across pages if files are added or removed while you are paging.
When saving the file, use the name in filename* (the UTF-8 name, percent-encoded) if it is present; otherwise use filename. For a name that contains characters such as Japanese, filename is a substitute name, and the original name is only in filename*. For example, for 売上.pdf ("sales.pdf" in Japanese), the header looks like this (the substitute name in filename is only an example).
Note: The Marketplace edition does not require scopes (authentication only).
GET/v1/search/messages
Full-text search messages. Messages of conversations of project steps and of archived conversations are not searched. Results are returned newest first by creation time (oldest first with order=asc), with cursor-based pagination.
Array of artifacts generated by the agent (included only when present)
messages[].deliverables[].filename
string
Artifact file name
messages[].deliverables[].filepath
string
Artifact file path
messages[].deliverables[].source
string
"agent" (agent-generated) or "user" (user-uploaded)
messages[].deliverables[].isPrimary
boolean
Whether it is a primary artifact
messages[].deliverables[].createdAt
string
Creation date and time (ISO 8601 format). An empty string when unknown
messages[].deliverables[].mimeType
string
MIME type (e.g., "application/pdf", "text/markdown"). "application/octet-stream" when it cannot be determined
messages[].deliverables[].size
number?
File size (bytes). Omitted when 0 or below
messages[].files
array?
Array of user-uploaded files. Included only when present for messages with sender: "User". The field structure matches the history API (fileId / mimeType / width? / height?). See GET /v1/conversations/:id/messages for details
hasMore
boolean
true when there is a next page. When true, cursor is returned
cursor
string
Cursor value for the next page. Included only when hasMore is true
limitReached
boolean
true only when many messages match and only the most relevant results (up to 1,000 each for matches in the message text and in the agent's task progress) are covered. Not included otherwise. To search exhaustively, narrow the query
q parameter is missing or not a single string (for example, q specified more than once), order is not "asc" / "desc", limit is not an integer from 1 to 100, or cursor cannot be read
501
An environment where search is not configured (code: "search_not_configured")
503
Search is temporarily unavailable (code: "service_unavailable"). Retry after a while
A project is the unit in which multiple agents are placed to carry out work. From an external system you can list projects, start a run, check run status, and get the artifacts of workflow runs.
There are two project types, and each is triggered differently.
Type
Description
How runs are triggered
workflow
Advances through steps in order. A run starts from the first enabled step.
POST /v1/projects/{projectId}/runs
cycle
Keeps multiple agents running on their own schedules. An external event can add one extra ad-hoc run.
POST /v1/projects/{projectId}/triggers/{triggerId}/fire
What you can operate on
With a service account token, you can operate on every project in this environment. An application (sign-in from an external tool) token can also call the nine endpoints here and the other project APIs. What it can operate on then depends on how the user participates in the project and on the organization operations role (User API Reference (Projects)). Calling them with a user token obtained in any other way returns 404 Not Found for every endpoint.
Retrieves the list of all projects in this environment. Supports cursor-based pagination. cycle projects also appear in the list, but they are not valid targets for starting a run (POST /v1/projects/{projectId}/runs).
The body can be omitted. If you send one, make it a JSON object. Fields other than those below are ignored.
パラメータ
型
必須
説明
payload
any
—
The request for the run. Any JSON value (an object, an array, a string, and so on) is accepted as is. The agent of the first step receives it in context.start_input (see "Request of a run started from outside" in the marketplace workflow guide). Putting a file ID in the request does not attach the file. Omitted or null means no request. Up to 8,000 characters as a JSON string
config
object
—
Input files and step settings. Write one entry per placement and step, in the form { "placements": [{ "placementId", "stepId", "parameters", "connectionBindings", "fileBindings" }] }. fileBindings maps input file names (inputs[].name from the list of registered Agents that can be placed) to the artifactId of uploaded drafts; parameters are the registered Agent's parameters (in the shape of parametersSchema); connectionBindings maps connection slot names to connector IDs. All fields other than placementId and stepId can be omitted. Omitted or null means no settings
There are two limits. If payload exceeds 8,000 characters as a JSON string, 400 Bad Request (payload_too_large) is returned. If the whole body exceeds 16 KiB, 413 Payload Too Large (body_too_large) is returned.
config also counts toward the 16 KiB limit of the whole body.
Resending with the same key returns the same run even after the project has been paused, after the concurrent-run limit has been reached, or after that run has finished.
If a timeout leaves the outcome unknown, simply resend with the same key. Nothing is run twice.
To start a different run, always issue a new key.
Resending with the same key but a different payload or config returns 409 Conflict (origin_conflict).
The Idempotency-Key length or character set is invalid
400
invalid_parameter
projectId is not a UUID
400
payload_too_large
payload exceeds 8,000 characters as a JSON string
400
invalid_body
The request body is malformed JSON or is not a JSON object
413
body_too_large
The request body exceeds 16 KiB
415
unsupported_media_type
A request body is sent with a Content-Type other than application/json
404
project_not_found
The project does not exist or is not accessible
400
approvers_required
An enabled review step has no reviewers configured
409
unsupported_project_type
A cycle project was specified
409
project_inactive
The project is not active (paused or archived)
409
principal_suspended
An administrator has suspended the user the project runs as. Ask an administrator to transfer the project to another owner, or to lift the suspension
409
start_step_in_branch_lane
The start step is in a lane of a branch step. Set the start step of the time trigger to a step outside the lanes of branch steps
409
definition_pending
The first definition of the project is waiting for approval, so it cannot run yet. Start it after an approver (another owner or an administrator of the organization) approves it in the product screen
409
relink_required
The project cannot be started or resumed until the references that were not found during import are reconnected. The remaining references appear in the project's relink and in the placements' relink. To clear the references that refuse a start: for rows whose kind is "agent", place the agent again (POST /v1/projects/{projectId}/placements); for rows whose kind is "connector", save the connector again on the product screen. Rows whose blocking is false are items to set up again if needed and do not refuse runs; they can be cleared with DELETE /v1/projects/{projectId}/relink, but clearing them alone does not make the project startable
409
admin_stopped
An administrator has stopped the project, or a registered agent placed in it. Ask an administrator to resume it (this API cannot resume it)
409
mixed_runtime
Registered agents and other agents are placed in the same project
409
agent_not_certified
A registered agent placed in the project is not certified for use in workflows
409
lanes_not_for_marketplace
The flow of the project has lanes (a project of registered agents cannot proceed with lanes)
400
invalid_parameter
A value in the start input is invalid
400
missing_input
An input file the step needs is missing (the previous step did not produce it, or it was not specified)
400
duplicate_target_path
Two or more input files would be placed at the same target path
400
mp_placement_not_in_snapshot
The registered agent of the step is not in the configuration taken when the run started
400
connection_version_missing
The connection version is not saved
400
invalid_runtime_config
The runtime settings of the placement do not match the registered agent definition
400
input_binding_cycle
The input bindings form a cycle
400
input_binding_forward_reference
An input refers to the output of a later step that has not run yet
400
input_binding_unknown_output
An output specified as an input does not exist in the source registered agent
400
artifact_format_mismatch
The format of an artifact specified as an input differs from the required format
400
artifact_digest_missing
An artifact specified as an input has no digest
400
invalid_config
config is malformed, names a placement or step that does not exist, does not match the registered Agent definition, or the project has no registered Agents placed on its steps
409
artifact_not_ready
An artifact specified as an input is not ready yet (or differs from the content that was reviewed). With a system integration token, only drafts that a current owner of the project uploaded and that are not yet bound to a run can be specified
409
origin_conflict
A different payload or config was sent with the same Idempotency-Key (the same origin was already used with different content)
409
connection_version_unavailable
The connection version cannot be confirmed (the connection target or permissions changed after approval)
429
too_many_active_runs
The project already has the maximum number of runs (3 in total) that are running or awaiting review
Fires a webhook trigger configured on a cycle project and runs the target agent once, as an extra ad-hoc run. Calling it for a workflow project returns 404 Not Found.
The scheduled runs stay exactly as configured. A fired run is added separately from the schedule. Each fire is accepted separately and is never merged into an earlier fire. Fires to the same step of the same agent placement are processed one at a time, in the order they were accepted.
A unique ID issued by the sender. Resending the same ID is treated as a duplicate. 1-160 characters using alphanumerics and _ . : @ + - only
payload
any
—
Arbitrary JSON passed to the agent. It is serialized to a JSON string and added to the request sent to the agent; anything beyond 8,000 characters after serialization is omitted
"accepted" (newly accepted) or "duplicate" (already accepted for the same eventId)
jobId
string
ID of the receiving job. It is fixed per combination of agent placement and step: fires to the same combination return the same value (it does not change per fire). It cannot be used to get run status (GET /v1/projects/{projectId}/runs/{runId})
executionId
string
ID of the accepted sweep execution. It can be used to get the result of the execution and to stop the sweep execution. A resend with the same eventId returns the same value (it can be missing from the response to a resend; then find the execution as described in Finding the Execution Started by a Fire). Until it starts (this can take about a minute, and longer when earlier fires to the same step of the same agent placement are waiting), getting the result of the execution returns 404 Not Found (execution_not_found)
This 202 Accepted means only that the request was recorded as accepted. The run happens afterwards, and its outcome is not part of this response. Do not treat 202 as business success.
Idempotency is decided by the combination of project, trigger, and eventId. Resending the same content returns 202 with result: "duplicate". Sending the same eventId with a different payload returns 409 Conflict (event_conflict).
The request body is malformed JSON or is not a JSON object
415
unsupported_media_type
A request body is sent with a Content-Type other than application/json
404
trigger_not_found
The project or trigger does not exist, is not accessible, the trigger is not a webhook trigger, or a workflow project was specified
409
trigger_disabled
The trigger is disabled
409
project_not_active
The project is not in a runnable state
409
principal_suspended
An administrator has suspended the user the project runs as. Ask an administrator to transfer the project to another owner, or to lift the suspension
409
admin_stopped
An administrator has stopped the project, or the Agent of the placement the trigger runs. Ask an administrator to resume it
409
placement_unavailable
The target agent placement does not exist, or is not in a state that can run
409
event_conflict
The same eventId was sent with a different payload
413
body_too_large
The request body exceeds 16 KiB
429
queue_full
The number of fires waiting to run on the same step of the same agent placement has reached the limit (20)
GET/v1/projects/{projectId}/runs/{runId}
Retrieves the status of a run. Besides runs started through the API (POST /v1/projects/{projectId}/runs), runs started from the UI or on a schedule can also be retrieved. cycle projects do not create runs, so they are out of scope for this endpoint.
Run state. One of "running" (running), "waiting_gate" (awaiting review), "closed_done" (completed), "closed_rejected" (ended by rejection), "closed_failed" (ended by failure) or "cancelled" (cancelled). Any value other than "running" and "waiting_gate" means the run has ended
originType
string | null
How the run started. One of "manual" (started from the UI), "api" (started through the API) or "cron" (scheduled)
originRef
string | null
A reference value that depends on how the run started. For a run started through the API, the Idempotency-Key used to start it
currentStepId
string | null
ID (UUID) of the step in progress. null for a completed run. For a run that ended for another reason, the ID of the step at that time may remain
startedByRef
string | null
Who or what started the run. See the "Who started the run" table below
actorLabel
string | null
Display name of a starter that is not a user (the service account name). null if unknown, and null otherwise
startedAt
string | null
Start time (ISO 8601). null before it starts
endedAt
string | null
End time (ISO 8601). null before it ends
summary
string | null
For a completed run, the output of the last completed step that recorded an output (first 2,000 characters). null if there is no such step and for a run that has not completed
runtimeKind
string | null
Kind of runtime. "marketplace" (a run of registered agents) or "autonomous". May be null for older runs
cancelRequestedAt
string | null
Time (ISO 8601) when cancellation was requested. null if not requested
deadlineAt
string | null
Deadline of the workflow run (ISO 8601). null if it has no deadline
configRevision
number | null
Revision of the case configuration. Used by runs of registered agents; null for other runs
steps
array
Records of the step runs, in the order they were recorded
steps[].runStepId
string
ID of the step run (starts with "rs:")
steps[].stepId
string | null
Step ID (UUID)
steps[].placementId
string | null
ID (UUID) of the placement that handled the step. null for rows that do not run an agent (such as review steps)
steps[].executionId
string | null
ID of the agent execution. Set once the agent starts running
steps[].conversationId
string | null
ID of the conversation of the agent execution. null for runs of registered agents
steps[].status
string
One of "pending" (waiting to run), "running" (running), "waiting_gate" (awaiting review), "done" (done), "failed" (failed) or "skipped" (skipped)
steps[].attempt
number | null
Retry count within the same visit (starting at 1)
steps[].lane
string | null
Always "main"
steps[].scheduledAt
string | null
When the step has a scheduled time, the time (ISO 8601) this row waits for. null once it starts
steps[].startedAt
string | null
Start time (ISO 8601)
steps[].endedAt
string | null
End time (ISO 8601)
steps[].outputSummary
string | null
Summary of the step output
steps[].outcome
string | null
Result of the agent execution. One of "completed", "incomplete", "failed", "timeout", "stopped", "denied" or "skipped". null for rows without a result
steps[].reasonCode
string | null
Reason code for the result. "branch_undecided" when a branch could not decide and a person is asked to choose; "branch_not_selected" for a step in a lane the branch did not choose
steps[].held
boolean
Whether the step is held because the project is paused and is waiting to resume
steps[].visit
number | null
Which visit to this step this is (increases with send-backs). null for rows without this record
steps[].reworkReason
string | null
Reason for the send-back
steps[].gateFor
string | null
For a review row, the step run (runStepId) it was raised for
steps[].routedFrom
string | null
When steps were skipped by choosing a destination, the step run (runStepId) that chose it
steps[].late
string | null
Why a result that arrived late was not used to advance the run
steps[].gate
object | null
Summary of the review. Set on review step rows. null for other rows
steps[].gate.decision
string | null
One of "proceed", "send_back" or "reject". null if not decided yet
steps[].gate.waitSource
string | null
What the review waits for. "human" (a person's decision) or "external" (an external condition)
steps[].resumeCommandId
string | null
Always null
steps[].resumeReason
string | null
Always null
steps[].supersededBy
string | null
Always null
steps[].gateKind
string | null
Always null
steps[].connectionConfirmOf
string | null
Always null
plannedPlacements
array
Order of steps and placements planned to run, based on the configuration when the run started
plannedPlacements[].stepId
string | null
Step ID (UUID)
plannedPlacements[].placementId
string | null
Placement ID (UUID). null for review steps and steps without an agent
The run does not exist, belongs to another project, or is not accessible
POST/v1/projects/{projectId}/runs/{runId}/cancel
Cancels a workflow run that is running or awaiting review (status is "running" or "waiting_gate"). This is the same as Cancel run on the screen.
When it is cancelled:
The status of the workflow run becomes "cancelled" (cancelled). This cannot be undone.
Steps that have not started are not started. The Agents of running steps are told to stop (the status of such a step row may not change until it stops).
Review requests of review steps that have not been decided are withdrawn.
A value to identify resends (8 to 128 characters of letters, digits, and _ . : -)
reason
string
—
The reason for canceling (up to 500 characters)
If you send it for a workflow run that is not running (after it ended or was cancelled), nothing changes: 200 OK is returned with existing set to true. The same applies when you resend after it was cancelled.
If you send it around the same time as another cancel or the end of the workflow run, the choice between 202 Accepted and 200 OK and the value of existing may not match what actually happened. Check the result with status of Get a Workflow Run (GET /v1/projects/{projectId}/runs/{runId}).
504 Gateway Timeout does not mean "it was not cancelled". Resend with the same commandId, or check status with Get a Workflow Run.
1curl -X POST https://<your-domain>/api/v1/projects/3f2a9c1e-5b7d-4e8a-9c2f-1a6b4d8e0f31/runs/7c1d5e2a-8f3b-4a6c-b9d0-2e4f6a8c0b13/cancel \
2-H "Authorization: Bearer <access_token>" \
3-H "Content-Type: application/json" \
4-d '{
5 "commandId": "cancel-20261118-0001",
6 "reason": "Because an error in the input data was noticed"
7}'
curl -X POST https://<your-domain>/api/v1/projects/3f2a9c1e-5b7d-4e8a-9c2f-1a6b4d8e0f31/runs/7c1d5e2a-8f3b-4a6c-b9d0-2e4f6a8c0b13/cancel \
-H "Authorization: Bearer <access_token>" \
-H "Content-Type: application/json" \
-d '{
"commandId": "cancel-20261118-0001",
"reason": "Because an error in the input data was noticed"
}'
import requests
response = requests.post(
"https://<your-domain>/api/v1/projects/3f2a9c1e-5b7d-4e8a-9c2f-1a6b4d8e0f31/runs/7c1d5e2a-8f3b-4a6c-b9d0-2e4f6a8c0b13/cancel",
headers={"Authorization": f"Bearer {token}"},
json={'commandId': 'cancel-20261118-0001', 'reason': 'Because an error in the input data was noticed'},
)
print(response.json())
const response = await fetch(
"https://<your-domain>/api/v1/projects/3f2a9c1e-5b7d-4e8a-9c2f-1a6b4d8e0f31/runs/7c1d5e2a-8f3b-4a6c-b9d0-2e4f6a8c0b13/cancel",
{
method: "POST",
headers: {
"Authorization": "Bearer " + accessToken,
"Content-Type": "application/json",
},
body: JSON.stringify({"commandId": "cancel-20261118-0001", "reason": "Because an error in the input data was noticed"}),
}
);
const data = await response.json();
console.log(data);
payload := []byte(`{"commandId": "cancel-20261118-0001", "reason": "Because an error in the input data was noticed"}`)
req, _ := http.NewRequest("POST", "https://<your-domain>/api/v1/projects/3f2a9c1e-5b7d-4e8a-9c2f-1a6b4d8e0f31/runs/7c1d5e2a-8f3b-4a6c-b9d0-2e4f6a8c0b13/cancel", bytes.NewBuffer(payload))
req.Header.Set("Content-Type", "application/json")
req.Header.Set("Authorization", "Bearer "+token)
resp, _ := http.DefaultClient.Do(req)
defer resp.Body.Close()
var result map[string]interface{}
json.NewDecoder(resp.Body).Decode(&result)
fmt.Println(result)
For a cycle project, stops a sweep execution. This is the same as Stop on the screen. No body is needed (a body is not used even if sent).
Only sweep executions of cycle projects can be stopped. Specifying an execution of a workflow step (rs:<number>) returns 400 Bad Request (invalid_parameter). To stop a workflow run, use Cancel a Workflow Run.
After it is accepted, it can take about a minute for the execution to stop. Check whether it stopped with Get the Result of an Execution (GET /v1/projects/{projectId}/executions/{executionId}). It can also be called with a system integration (service account) token (it returns only the status fields). An execution that had not started yet when the stop request was accepted stays 404 Not Found (execution_not_found) in Get the Result of an Execution.
For an execution whose stop was already requested, or that had already ended, nothing changes: 200 OK is returned with existing set to true.
202 Accepted when a new stop request was accepted, and 200 OK otherwise.
Field
Type
Description
executionId
string
ID of the execution
state
string
The state when the request was accepted. It does not indicate whether the execution actually stopped (see the first bullet of this section for how to check). Example values: "dispatched" (the stop request was passed to the running execution), "cancelled" (an execution that had not started yet was cancelled), "terminal" (the execution had already finished). Other state values of the execution or the dispatch at the time of acceptance may also be returned as they are. Do not fix a list of values and rely on it
existing
boolean
true when a stop had already been requested or the execution had already ended (always false with 202 Accepted)
Gets the status of an execution (whether it is running, whether it has ended, and how it ended). With a service account token, it returns only the status fields below. Called with an application (sign-in from an external tool) token, it returns the same fields as Get the Result of an Execution in the User API Reference (Projects).
A sweep execution of a cycle project can be read with executionId in the response of a trigger fire. Until it starts (this can take about a minute), it returns 404 Not Found (execution_not_found). Use it also to check whether a sweep execution stopped after you stopped it.
An execution of a workflow step can be read with steps[].executionId (a UUID) of the workflow run. executionId in the response is rs:<number>, and an ID in this form can also be used.
The fields returned when called with a service account token.
Field
Type
Description
executionId
string
Execution ID. rs:<number> for a workflow, and a UUID for a cycle
kind
string | null
"run_step" (a workflow step) or "sweep" (a cycle sweep)
runId
string | null
ID of the workflow run. null for a cycle
stepId
string | null
Step ID
placementId
string | null
Placement ID
attempt
number | null
Attempt number (increases when a platform failure is retried. For a cycle, a relaunch is a new execution, and it does not increase)
outcome
string | null
Outcome (see Outcome Values in the User API Reference (Projects))
reasonCode
string | null
Reason for the outcome. For a cycle, the reason the execution ended (for example, "guardrail_blocked"); for a workflow, the reason reported by the step
resultState
string | null
"terminal" (has a result), "running" (running), "missing" (insufficient record), "unmeasured" (completion not measured), or "not_applicable" (a row that is not an Agent execution, such as a review step)
Gets all artifacts of a workflow run (the input files bound at the start and the outputs made by the steps), in the order they were made. Read the content with the API that gets the content of an artifact (GET /v1/projects/{projectId}/artifacts/{artifactId}/content). With a service account token, you can read the artifacts of every project in this environment. run_id is required.
Gets the content of an artifact. You can also call it with HEAD. The same headers are returned without the body, so you can check the size first. Reading it with GET leaves a view (forensic.viewed) in the records (audit trail) (HEAD does not). When read with a service account token, it is recorded as the system integration.
The same 404 Not Found is returned whether the project does not exist or is not accessible, and the response does not reveal which of the two applies. The code is fixed per endpoint: project_not_found for starting a run, trigger_not_found for firing a trigger, and run_not_found for run status and for canceling a workflow run. Listing artifacts returns project_not_found (run_not_found when the run does not exist), and getting the content of an artifact returns artifact_not_found. Stopping a sweep execution and getting the status of an execution return execution_not_found.
When the type is out of scope
Calling trigger fire on a workflow project, or run status or canceling a workflow run on a cycle project, returns 404. Stopping a sweep execution on a workflow project also returns 404. Starting a run on a cycle project is the one exception: it returns 409 unsupported_project_type.
Request body
When you send a body, make it a JSON object. Anything other than Content-Type: application/json returns 415, malformed JSON or a body that is not a JSON object returns 400 (invalid_body), and a body larger than 16 KiB returns 413.
Unknown fields / queries
Fields and query parameters that are not in the specification are ignored. An invalid value for a documented query parameter, or a cursor that cannot be read, returns 400 Bad Request.
Timeouts
504 Gateway Timeout does not mean "it did not run". Resend a run start with the same Idempotency-Key, a trigger fire with the same eventId and the same body, and a workflow run cancel with the same commandId. A sweep execution stop can also be resent as it is.
HTTP status codes that may be returned by all endpoints.
Status
Description
400 Bad Request
Invalid request parameters
401 Unauthorized
Token is invalid or missing. Also returned for a token of a user whose use was suspended by an administrator (see "Suspended users" below)
403 Forbidden
No access permission to the resource
404 Not Found
Resource does not exist, or the path or method is not defined (405 is not returned)
409 Conflict
Operation conflict, such as the resource currently being processed
409 Conflict
A person with the organization operations role (the "Operator" role assigned in "Admin Roles" in the admin console) who does not take part in the project called an operation that role cannot use (error.code is forbidden_role). It does not clear by waiting. For an owner-only operation, ask an owner of the project to do it. For other operations, ask an owner to add you to the project; you can then call them as your project role allows
413 Payload Too Large
Request body exceeds the size limit
415 Unsupported Media Type
The request body format (Content-Type) is not supported
429 Too Many Requests
Rate limit or concurrent-run limit exceeded. error.type is rate_limit_error, and error.code depends on the endpoint (for example monthly_quota_exceeded for the monthly limit and too_many_active_runs for concurrent project runs). Please wait and retry.
500 Internal Server Error
Internal server error
502 Bad Gateway
Communication error with upstream services (authentication / chat backends, etc.)
The parameter that caused the error. Included only for validation errors
error.reason
string
The kind of reason for the refusal. Included only for 409 (definition_invalid) of the project APIs
error.suggested_action
string
Recommended action to take. Included only for some errors
JSON
{ "error":{ "type":"invalid_request_error", "message":"Invalid parameter: limit", "code":"invalid_parameter", "param":"limit", "suggested_action":"limit must be between 1 and 100" } }
A call with the token of a user whose use was suspended by an administrator returns 401 Unauthorized. error.code is account_suspended.
The following APIs do not check whether the user is suspended: GET /v1/organizations, GET /v1/job-types, GET /v1/user, PATCH /v1/user, GET /v1/marketplace-agents, the MCP tools list_agents and cancel_chat (when no task is in progress), and the Agent Card of a registered Agent (GET /api/v1/a2a/agents/{agentId}/.well-known/agent-card.json).