model-context.md

Model context

Open Settings → Model context to set a model's working window and the token count that starts automatic compaction. On phones, tap the chat's context indicator to open this menu directly. On larger screens, Adjust limits in the chat's context view opens it. Changes belong to the selected server and configuration profile; the header identifies that server.

Set the window and compaction threshold

Choose Models or Families, then find the target by name, ID, or provider. Enter the Context window and Auto-compact after values in tokens and save. For example, a model supporting a million-token window can use 1000000 with automatic compaction after 500000. Numbers can be pasted with ordinary thousands separators. Decimal, negative, and unsafe integer values are rejected.

The window and threshold are independent. Lowering the threshold starts compaction earlier without shrinking the request budget available to compact the existing history. The automatic-compaction switch can disable the trigger for a target while retaining its saved numbers. Manual compaction remains available.

Blank fields inherit the existing defaults. Reset to defaults clears the target's personal rule when saved. A family rule overrides configuration defaults; a base-model rule overrides its family, and an exact alias rule overrides its base. Fields inherit independently. A family change preserves exact personal exceptions and shows the effective range plus the individual models beneath it. Identical family names at different providers remain separate.

Effective limits shows the window and threshold that can actually apply. A requested value cannot bypass the provider input limit, ChatGPT catalog ceiling, or the space needed for an answer. For an account advertising an 872000 ChatGPT working maximum, requesting 1000000 still applies 872000; an independent 520000 compaction threshold can fit that window. A window with no room for a response is rejected. Models without a known context limit keep their ordinary behavior unless a personal window is supplied.

On phones, the model list and editor use separate views. The two numeric fields, effective limits, and Save action have priority. Save occupies its own row, and the editor scrolls inside the visible viewport when the keyboard reduces it. Labels use compact 13 px text on phones and 14 px on larger screens. Numeric inputs retain 16 px text and touch controls retain their tap areas. Wide screens show the target list and editor together. Unsaved edits are protected when navigating or closing the menu.

When changes apply

A save does not restart the server, dispose projects, reconnect providers, or stop tools. The active logical turn keeps its captured settings through tool calls, retries, compaction, and persisted continuation. The next turn receives the saved values. Clients refresh through their existing event stream and when the menu reconnects or becomes visible.

The chat's context indicator uses the window from its most recent model request. Saving a new window while the chat is idle therefore changes that indicator on the next request. A smaller ChatGPT window can trigger a transition that first compacts history with the previous working window, including when the model ID stays the same.

Compaction is checked between model requests. Its count includes the active history, instructions, cached input, and the normal provider-specific accounting for tool results and reasoning. It is not a stop at an exact individual token. ChatGPT retains its server-side compaction; other providers retain OpenCode's text-summary mechanism. The menu changes their limits, not their summary format.

Existing configuration remains a fallback. Personal UI rules take precedence without rewriting opencode.jsonc. Requests from Web, TUI, and API clients using that server/profile receive the same settings.

Import and export

Export settings opens a selection of saved personal model and family rules. Download the selected rules as JSON. The bundle contains requested values, automatic-compaction choices, and portable target identifiers. It contains no credentials, server addresses, filesystem paths, or conversation data. Effective limits are resolved again at the destination.

Import settings accepts an opencodez-context JSON bundle, version 1. Review the preview before applying it. New targets are selected by default; unchanged targets are already present, and differing local rules are kept until you explicitly choose their replacement. Re-importing an exported bundle creates no duplicates. Unknown models and families can be retained as inactive rules. They become applicable when the target is available in the profile.

Imports allow at most 1,000 rules and 1 MiB. Duplicate targets, invalid values, unsupported formats, and unusable known windows leave stored settings intact. Saving checks the document revision. A conflicting editor keeps its draft and offers refresh; refreshing an import reevaluates its replacements.

Storage and maintenance

Personal rules live in models/context.json beneath the active OpenCodez config root. The file is user data: include it in profile backups and preserve it during application updates and configuration reconciliation. Writes replace the file atomically. No service, database dependency, recursive watcher, or polling loop is added. Editing a value previews only the affected models, using the existing catalog cache rather than rescanning conversation history.

The browser-safe contract is in schema/src/opencodez-context.ts. Rule resolution and captured values belong to core/src/opencodez/context-policy.ts; storage, portable bundles, and effective limits belong to opencode/src/opencodez/context-settings.ts. The existing OpenCodez HTTP group owns the endpoints, and both Settings layouts mount the same app/src/opencodez/context-settings.tsx screen only while its tab is active. Sampling and compaction reuse the captured model limits. Keep public API changes and their generated SDK/client output together.