Task scope and result reads¶
Identity and tool selection¶
These two task paths use the current request identity, workspace, and task-saved tool list. The backend rechecks current agent-use permission and resolves private connection credentials for the current caller, without using the task owner to obtain credentials for another member. Completed tasks still support app reads and calls; failed, canceled, and rejected tasks do not. Terminal states are neither universally allowed nor universally prohibited.
Resource reads match URI and optional tool_id against the task list. URIs must have no leading/trailing whitespace or userinfo. They are absolute resource identifiers, not an entry point for arbitrary browser-supplied server URLs. If multiple tools share a URI and no ID is specified, the current implementation chooses the first match rather than rejecting every ambiguity.
tool_name takes precedence over tool_id. Names may match platform names or persisted original MCP names/service names. Multiple matching tools allowed for app access are rejected. A valid app description is required. Unrestricted visibility permits both uses; an explicitly model-only tool cannot be invoked by the app. Leading/trailing whitespace in IDs/names is rejected.
Resource and invocation results¶
Response location |
Reading rule |
|---|---|
|
Resource content uses camelCase. Do not rename it to mime_type. |
|
The MCP read branch requires exactly one of the two. blob is valid Base64; empty text still counts as supplied text. |
|
Product invocation results use snake_case and may contain any JSON value or explicit null. |
|
Ordinary false is also returned. true indicates tool-result failure; HTTP 200 does not establish business success. |
The MCP resource Provider validates 1–16 content blocks. Each text or Base64-encoded string is at most 2 MiB. This is a per-block limit, not a whole-response cap. Published API service pages use the corresponding workflow resource owner; not all MCP network-branch validation applies. Resource block URIs are only checked for nonemptiness and need not equal the request URI.
MCP tool calls preserve provider isError as product is_error. Handle network/protocol failures separately from tool-reported failures. Invalid provider results in current resource reads may return 502. Provider exceptions in calls map through runtime unavailability to 503; the same HTTP table does not apply to both.
Invocation results return directly to the caller and do not establish updates to model conversations or the entire run result. Real tools may write or make external calls. After failure/timeout, query the tool’s business result before retrying. Task completed does not mean tool success. Published API service calls enter their owner with the current caller and effective role; browsers cannot supply role fields to override this.