Skip to content
OneviumDocs
On this page

Configure and verify a 1M context window

Confirm model and account support, configure the global switch and model capability, then verify the actual window in a new conversation.

Confirm support before enabling a larger window#

A 1M context window helps when a request needs to include more history or reference material. It does not enlarge every model's capacity, guarantee better answers, or keep costs unchanged.

Onevium needs to know both whether this model on this connection supports 1M and whether you want to enable it. From 1.2.2, OpenAI Responses, Chat Completions, and Gemini connections can combine model capabilities with the global switch. You normally do not need a duplicate model entry with a [1m] suffix.

Before you start#

  • Confirm with your provider that the specific model, endpoint, and account plan support at least 1M tokens. The same model name behind another gateway may have different limits.
  • Get a successful short reply first. See Providers and models and Custom providers.
  • Check the provider's Connection type. Do not change protocols to enable a larger window.

Settings for each connection type#

Connection typeSettings to check
Official Anthropic / Claude Code loginUse a model available to your account; check the global 1M Context Window in Claude Code settings when you need its long-context option
Custom Anthropic Messages endpointEdit provider → Advanced options → Endpoint supports 1M context, plus the global 1M Context Window; enable only after confirming support
OpenAI Responses, Chat Completions, Google GeminiManage models → the model's Advanced fields → Confirm 1M support for this model, plus the global 1M Context Window
Dedicated long-context entries in compatible-service presetsPreserve the preset mapping and check the model and plan; an (1M) display name does not grant account access

Built-in capability data can identify models on known official endpoints. Custom gateways and unrecognized models require your explicit confirmation after checking support. Entering a large window number cannot make the endpoint support it.

Configure an OpenAI or Gemini connection#

  1. Open Settings → AI services, find the connected service, and open Manage Models.
  2. Find the model, confirm it is enabled, and expand its Advanced fields.
  3. If the app cannot identify support but your provider has confirmed it, enable Confirm 1M support for this model. Leave it off for models or plans below 1M.
  4. Click Save, then enable the global 1M Context Window in Claude Code settings.
  5. Create a test conversation, select that provider and model, and verify it below.

Keep the real identifiers required by your connection in Model ID and Upstream model ID. The current workflow handles the internal long-context marker where applicable. Do not add [1m] to the provider's upstream model ID.

If an older setup has a manually created [1m] duplicate, first verify the original model using this workflow. Then decide whether to disable the duplicate. Do not delete an entry still used by conversations or automations just to tidy the menu.

Check role mappings and subtasks#

The main conversation's model, Claude Code role mappings, and AI-created child conversations are related but separate choices.

  • Role mappings: Edit the provider and check Fable, Opus, Sonnet, Haiku, and Subagent in the advanced model mapping fields. In 1.2.2, OpenAI Responses, Chat Completions, and Gemini connections also support these mappings. Choices come from enabled models on the same provider. Save a new connection and configure its models before returning to set mappings.
  • Each model's capability: Mapping a role does not transfer 1M support to another model. A role assigned to a smaller-window model follows that model's capabilities.
  • Actual child conversations: Open the member after creation and check its selected model and window. A successful main-conversation check does not establish the budget of every member.

Anthropic-compatible connections have their own role fields and environment overrides. See Custom providers.

Verify the window and a real reply#

  1. Create a new conversation, select the target model, and send /context to inspect the actual total window.
  2. For a model configured through this guide, confirm that the session reflects the long-context setting. Do not rely only on the number shown in Manage models.
  3. Send “Reply only: received” and check that the model responds successfully.
  4. If you use role mappings or child conversations, check their actual models and /context results separately.

/context checks the budget used by the conversation; a short reply checks a basic request. Neither alone proves that the provider accepts a request near the limit. Real tasks remain subject to account limits, rate limits, tool capabilities, and server constraints.

Troubleshooting#

What you seeWhat to check
The window still shows roughly 200KThe global 1M switch, the selected model's support status, and whether the new test conversation uses the intended entry
The official model works but the same name on a gateway does notEndpoints can differ; confirm gateway support, then configure its own model entry
The provider reports an unknown modelThe upstream ID, an accidental [1m] suffix, and account access to that model
The main conversation works but a subtask failsThe role must map to an available, enabled model on the same provider
Per-model confirmation or native-protocol role mapping is missingCheck your installed version; these steps require 1.2.2 or later
Changing the connection type causes protocol errorsRestore the endpoint's supported protocol; long context does not require a protocol change

Next steps#

Use Providers and models to manage available entries or Custom providers for a team gateway. Set default reasoning effort in the relevant provider's model configuration.