Conversation history and state
Keep conversation history in your application or client. A request ID in Primex Usage correlates a request and charge; it is not a conversation ID or a copy of the response. Each new model turn has its own usage and billing.
Choose the continuation format
| API path | History your application sends |
|---|---|
| Chat Completions | A messages array with the relevant user, assistant and tool messages |
| Native Responses | Supported input items, including prior output/tool items when replaying history |
| Messages | The relevant user/assistant content blocks, with system sent separately |
| Responses rewritten to Chat | A complete input/history that can be represented as Chat messages; native Responses state is not provided |
For a simple Chat conversation, append the returned assistant message, then the
next user message to your stored messages. Do not send only the latest question
if it depends on earlier turns. Keep enough context for the task while respecting
the model's context and current rate limits.
Tool-result continuations
Preserve the model's tool call and its identifier before executing the tool. Return the result in the same API format:
- Chat Completions: append the assistant's
tool_calls, then arole: "tool"message with the matchingtool_call_id. - Native Responses: retain the required previous output items and send the
matching
function_call_outputusing the originalcall_id. Custom tools use their own supported output item shape. - Messages: retain the assistant's
tool_useblocks, then supply matchingtool_resultblocks withtool_use_idin the next user turn.
Send a result for each tool your application executes. Do not invent new IDs or display tool calls as a completed text answer. See the runnable function-tool loop.
Native Responses state
Native Responses forwards supported provider-native request fields. State features
such as previous_response_id, store and encrypted reasoning depend on the
actual model/provider path. Native Responses support alone does not establish
that all state features are available.
Primex does not publish a universal conversation-store lifetime, a general
Responses retrieval/deletion API, or a guarantee that a saved provider response
can continue across a different model or serving account. Keep the history you
need for your application; do not use a provider response ID as your only backup.
store: false is a provider request option where supported, not a statement about
all Primex logs or data retention. See Privacy and team evaluation.
When replaying native output, preserve required reasoning and tool items according to the selected route's contract. Encrypted reasoning/compaction items are opaque: do not modify them or assume they are portable across models/providers.
The separate POST /v1/responses/compact route requires a native Responses model.
Compaction is provider model work and can be billed; it is not a free metadata
read or a Chat Completions feature. The provider's supported compaction contract
applies. Chat-only models return unsupported_responses_compaction.
Files and hosted containers
Use Primex Files with the correct purpose, ownership and expiry. A temporary file reference does not preserve the surrounding conversation. Claude Files and hosted Skills have a separate contract; the documented Skills workflow requires a fresh container and does not allow container reuse. Generic Messages history does not enable those features.
Recover a failed continuation
For invalid_encrypted_content, preserve the safe error code and request ID.
Start a fresh client session or deliberately rebuild supported plain conversation
history with the user. Do not automatically strip opaque items, retry the same
payload, or replay tool side effects to conceal a state error.
After a timeout or interrupted stream, inspect Usage before starting another paid turn. Retain already-executed tool outcomes so a recovery does not repeat external actions. See Errors & Limits and Streaming.