dsh-operating-context

When running multiple model services under DeepSeek Harness, context window sizes are scattered across the catalog, per-model lists, and defaultContextWindow, with no intuitive single place to keep them consistent. Hand-editing settings.yaml is error-prone, and patching adapter config directly risks wiping endpoints, credentials, and other critical fields.

dsh-operating-context adds a Working Window option to the Settings page. All writes go through settings.mutate in the user layer: it does not mount a second compaction engine or alter adapter internals. It reads each model's native ceiling from the installed catalog, guaranteeing the chosen window never exceeds what a model can actually hold. When the selected value is at or below a model's ceiling, the plugin removes any existing capacity override so the catalog native value takes effect again, making the operation idempotent and reversible. Cross-namespace batch writes report the exact number of completed batches on partial failure and reload authoritative state rather than presenting a partial write as a rollback.

Ideal for users managing multiple providers (DeepSeek, Kimi, Anthropic, OpenCode, and beyond) in Harness who need a unified context window size and want to avoid manual YAML edits. Catalog routes, routes with explicit model lists, and hand-declared routes are all supported through the same settings page.