roboto.api_version

Release date for v0.35.0 of the Roboto Python SDK

Content mode introduced for query APIs.

path_in_schema is now a required field on AddMessagePathRequest

Query endpoints are now eventually consistent when using v2026_03_13 or later. Clients on older API versions maintain strong consistency for backward compatibility.

AgentSession and AI Chat are renamed to AgentThread across the SDK and REST API: chat_id / session_id become thread_id on the wire, /v1/ai/chats becomes /v1/ai/threads, and agent invoke becomes launch (POST /v1/ai/agents/{id}/launch; the previous developer-only …/invoke URL was removed outright). Clients on older API versions continue to receive session_id in response bodies and can keep calling the legacy /v1/ai/chats paths, which are soft-deprecated aliases on the same handlers.

PUT /v1/files/upload/<id>/progress responds with {uri, file_id} pairs instead of a bare file-id list, so upload clients can tell which created file belongs to which uploaded path. Clients on older API versions continue to receive the bare id list.

GET /v1/datasets/tags, GET /v1/events/tags, and GET /v1/files/tags respond with a paginated {items, next_token} object (with optional search/limit/page_token query parameters) instead of a bare string list, so tag autocomplete scales past the unique-tag cap. Clients on older API versions continue to receive the bare list.

/v1/triggers speaks the v2 trigger shape (events/once_per/targets from roboto.domain.triggers) instead of the legacy causes/for_each/action-column shape, and rejects a caller-supplied service_user_id. Clients on older API versions keep the legacy shape: their creates and updates are translated onto v2 triggers, responses are projected back, and triggers with no legacy representation (multiple targets, non-action targets, new event types) are filtered from lists and 404 on direct reads.

Agent thread messages can carry client_context content blocks: what the user was viewing when they sent the message, stored on the message itself. Clients on older API versions receive threads with those blocks removed, since their AgentContent union cannot parse them.

A file record can be a link: fs_type "link" and a roboto:// uri pinning one version of another file. Clients on older API versions, whose FSType has only file and directory, never receive one: links are dropped from file listings, and reading one directly by ID or path reports it as not found.

POST /v2/topics/id/<topic_id>/read-plan states the bounds each partition’s rows are selected in on a per-partition window, instead of the extent it returned before. Clients on older API versions continue to receive extent, set on every partition to the plan’s own window; such a client selects rows by the plan’s window alone. Two kinds of read are refused with HTTP 400 and a message to upgrade rather than served that way: a read scoped to a Session that holds a file over only part of the requested time span, and a read covering a file that packs several partitions’ data. Either would otherwise return rows belonging to another Session, or to another partition packed into the same file.

Adding files to a Session, removing them, and publishing metrics answer with one element per entry the request named, in the order it named them, each holding the entry’s result or the error that refused it. An entry the platform refuses leaves the others in place: an add decides every refusal before writing anything and adds the rest together, and a remove removes together every file the Session holds and reports the rest as not held. A failure the platform did not anticipate, such as a timeout, adds or removes none of them. An add naming a file that does not exist is refused whole, before any entry is applied. A file entry states the window the Session holds the file over in the file’s own timestamps, on min_file_timestamp_ns and max_file_timestamp_ns, and a call names at least one file. Listing a Session’s files reports each file’s window in Unix-epoch nanoseconds on min_wall_clock_timestamp_ns and max_wall_clock_timestamp_ns.

Clients on older API versions keep the earlier contract for these four calls. Listing a Session’s files reports the same window on range_min_timestamp_ns and range_max_timestamp_ns. A file entry carries the file’s id, an optional data_range, and a window in Unix-epoch nanoseconds on range_min_timestamp_ns and range_max_timestamp_ns. Adds and removes accept a call naming no files and answer with the Session’s refreshed record. An add stands or falls whole: when the platform refuses a file, it answers with the error for the first refused file and adds none of them, and a file named twice is added over its last entry. A remove stands or falls whole too, and skips a file the Session does not hold. Metric publication answers with a succeeded list and a failed list. succeeded holds one record per metric and Session, carrying the last value written when a call published the same pair more than once; each entry in failed is named by the metric name the caller submitted.