chat_with_agent
ChatGPTSends a message to a Prime Agent (SEND), reads a run's progress (POLL), or answers a run's pending question (HITL_RESPONSE). Returns status (RUNNING, HITL_REQUIRED, COMPLETED, FAILED or CANCELLED), messages, artifacts, citations, poe, session_url, execution_handle, next_poll_handle (the input to poll_agent_run, a read-only progress read) and next, which names the call that continues the run. A .zip attachment is unpacked into its member files. Knowledge-base sources appear in message content as knowledge:// links and under citations; get_knowledge_document resolves a citation's uri, and the same uri is readable as an MCP resource. A file the content links appears as artifact://<artifact_id>, the file_handle get_artifact reads. Hosts that support MCP elicitation are asked inline while the call is open. Otherwise status HITL_REQUIRED carries interaction: mode FORM has response_schema and allowed_actions, answered by HITL_RESPONSE with interaction_handle, an allowed action and an instance of response_schema as interaction_response (APPROVE/REJECT take the same object, CANCEL none); mode URL has deep_link, where the user answers in Prime; mode WAIT names waiting_on, the person who decides. session_url, when present, opens this conversation in the Prime console and is the same on every call for one session. session_id is an opaque continuation handle that changes on every response.
chat_with_agent_team
ChatGPTSends a message to a Prime Agent Team (SEND), reads a Team Room turn's progress (POLL), or answers the decision a turn waits on (HITL_RESPONSE). Returns main-room status, the accepted synthesized answer, citations (knowledge:// sources, resolvable with get_knowledge_document or as MCP resources), session_url, execution_handle, next_poll_handle (the input to poll_agent_team_turn, a read-only progress read) and next, which names the call that continues the turn. Member messages, task output, tool results and collaboration traces are not returned (get_session returns the owner's view of the main room, members' posts included). A .zip attachment is unpacked into its member files. Status HITL_REQUIRED carries the interaction the turn is blocked on (prompt, interaction_handle, allowed_actions); the turn does not advance until it is answered. session_url, when present, opens this conversation in the Prime console and is the same on every call for one session. session_id is an opaque continuation handle that changes on every response.
get_agent_detail
ChatGPTGet public details for one Prime Agent selected from list_agents, including knowledge_bases: each knowledge base your runs of the Agent get, by dbName, with accessMode READ_ONLY or READ_WRITE (an Agent saves into a knowledge base only when it is READ_WRITE).
get_agent_team_detail
ChatGPTGet public details for one Prime Agent Team selected from list_agent_teams.
get_artifact
ChatGPTResolve one caller-authorized artifact: its verified metadata and a short-lived, self-authorizing download URL (a bearer capability) for the whole file. With include_content=true the result also carries one page of the bytes inline: text for text-like files (markdown, html, csv, json, svg...), base64 otherwise; at most 32768 bytes of text or 24576 bytes of other files per page, starting at offset. content_total_bytes is the file's size; next_offset, present while the file continues, is the offset of the next page.
get_knowledge_document
ChatGPTResolve a knowledge-base citation exactly as an answer cites it (knowledge://<db>/<document>[@v<n>][#chunk=<chunk>]), at the cited version (a version the document never had is KNOWLEDGE_DOCUMENT_NOT_FOUND): the document's name, type and size, a short-lived download URL for its original bytes, the cited chunk's text when a chunk is named, and with include_content=true the parsed text of the document in pages of up to 40000 characters. content_total_chars is the whole text's length; next_content_cursor, present while the text continues, reads the next page as content_cursor. The same URI is also readable as an MCP resource.
get_poe_report
ChatGPTFetch the Proof of Execution (PoE) certificate of one of the caller's Agent runs by its poe_ref, the EAC-... serial returned in a chat response's poe object. The ref does not expire, so it can be stored with the artifacts a run produced and fetched any time later. Returns status (GENERATING, READY or FAILED), the four hashes, the signature and, once READY, the signed payload including artifactManifest: the artifact ids and sha256 of the files the run attached, used and generated. When artifactManifest.complete is false, manifestGaps says why, as codes with counts and descriptions; outcome says whether the attested answer held (SEMANTIC_FAILURE when the agent could not stand behind it).
get_session
ChatGPTGet one page of a caller-owned Prime session selected by opaque session_id: the newest messages in time order, at most limit of them and about 40,000 characters in all. Each message carries its attachments (canonical artifacts, readable with get_artifact) and its citations (knowledge:// sources, readable with get_knowledge_document); a file the text links appears as artifact://<artifact_id>, the file_handle get_artifact reads; a message longer than 30,000 characters is cut and has content_truncated true. While older messages remain, has_more is true and next_cursor reads the next older page with the same session_id and limit.
get_session_message
ChatGPTRead the full text of one public message returned by get_session, in bounded pages. Pass the same session_id and the message_handle. offset counts Unicode characters, starting at zero; max_chars is 1 through 16000, default 16000. Append content pages in next_offset order until content_truncated is false, then parse a JSON answer. Every page carries the SHA-256 of the full text as content_sha256; restart at zero if that hash changes. A message handle expires with the session handle lifetime.
list_agent_teams
ChatGPTList the Prime Agent Teams available to the authenticated caller. Each team_id is stable for this user across calls and new Team versions, so it can be reused without listing again. Results are paged (20 by default): while next_cursor is not null, more Teams exist, and next_cursor as cursor with the same query reads the next page.
list_agents
ChatGPTList the Prime Agents available to the authenticated caller. Each agent_id is stable for this user across calls and new Agent versions, so it can be reused without listing again. Results are paged (20 by default): while next_cursor is not null, more Agents exist, and next_cursor as cursor with the same query reads the next page.
list_session_artifacts
ChatGPTList every canonical artifact of one caller-owned session, newest first: files the user attached and files the Agent, or a Team member posting into the main room, produced; at most 100 entries, with truncated true when more exist. An entry's file_handle is the artifact_id get_artifact takes for a download URL or, with include_content=true, the bytes inline.
list_session_runs
ChatGPTList durable Agent runs or Team main-room turns in one caller-owned Prime session. The next_cursor is opaque and reads the next page with the same session_id and limit.
list_sessions
ChatGPTList active Prime sessions owned by the authenticated caller.
list_uploads
ChatGPTList the authenticated caller's Knowledge artifacts with bounded pagination. status=AVAILABLE (default) lists published artifacts; status=LEASED lists the caller's reserved uploads that can still be finalized, each with lease_expires_at.
poll_agent_run
ChatGPTReads the progress of a Prime Agent run started by chat_with_agent, given its latest next_poll_handle (or execution_handle) as execution_handle. Returns the same response as chat_with_agent, including next. Read-only: it does not answer a pending interaction (that is chat_with_agent HITL_RESPONSE).
poll_agent_team_turn
ChatGPTReads the progress of a Team Room turn started by chat_with_agent_team, given its latest next_poll_handle (or execution_handle) as execution_handle. Returns the same response as chat_with_agent_team, including next. Read-only: it does not answer the room's pending decision (that is chat_with_agent_team HITL_RESPONSE).
recover_agent_run
ChatGPTRecover an existing Agent run or Team main-room turn after its execution handle expired. This only returns a new short-lived execution_handle for the selected run; it never sends a new message or creates a run.
update_session
ChatGPTSets the title of one caller-owned Prime session to the given text.
upload_file
ChatGPTUpload one inline file to Knowledge and return its verified artifact descriptor. Prime reserves the upload, sends the bytes and finalizes it in this one call; the byte transfer does not carry the user's bearer token. An idempotent retry of an already AVAILABLE artifact skips the PUT and finalize calls. The response contains only the canonical artifact_id/file_handle and verified metadata, never a storage URL. The decoded file is limited to 5242880 bytes.